Defined-Trust Transport (DeftT) Protocol for Limited Domains

1.1. Environment and use

Due to physical deployment constraints and the high cost of wiring, many OT networks preferentially use radio as their communication medium. Use of wires is impossible in many installations (untethered Things, adding connected devices to home and infrastructure networks, vehicular uses, etc.). Wiring costs far exceed the cost of current System-on-Chip Wi-Fi IoT devices and the cost differential is increasing [WSEN][COST]. For example, the popular ESP32 is a 32bit/320KB SRAM RISC with 60 analog and digital I/O channels plus complete 802.11b/g/n and bluetooth radios on a 5mm die that consumes 70uW in normal operation. It currently costs $0.13 in small quantities while the estimated cost of pulling cable to retrofit nuclear power plants is presently $2000/ft [NPPI].¶

Many OT networks are Limited Domains having a defined membership and communications that are often local, have a many-to-many pattern, and use application-specific identifiers ("topics") for rendezvous. This fits the generic Publish/Subscribe communications model ("pub/sub") and, as table 1 in [PRAG] shows, nine of the eleven most widely used IoT protocols use a topic-based pub/sub transport. Pub/sub protocols communicate by using the same topic but need no knowledge of one another. Today, these protocols are typically implemented as an application layer protocol over a two-party Internet transports like TCP or TLS which require in-advance configuration of peer addresses and credentials at each endpoint and incur unnecessary communications overhead Section 1.2. For example MQTT, an open standard developed in 1999 to monitor oil pipelines over satellite [MQTT][MHST], is now likely the most widely used application communication protocol in IoT (https://mqtt.org/use-cases/). Microsoft Azure, Amazon AWS, Google Cloud, and Cloudflare all offer hosted MQTT brokers for collecting and connecting sensor and control data in addition to providing local pub/sub in buildings, factories and homes.¶

1.2. Transporting information

Figure 3 shows a smart lighting example with a topic-based pub/sub application layer protocol in a wireless broadcast subnet. Each switch is set up to do triple-duty: one click of its on/off paddle controls some particular light(s), two clicks control all the lights in the room, and three clicks control all available lights (five kitchen plus the four den ceiling). Thus a switch button push may require a message to as many as nine light devices. On a broadcast transmission medium each packet sent by the switch is heard by all nine devices. IPv6 link-level multicast provides a network layer that can take advantage of this but current IP transport protocols cannot. Instead, each light switch needs to establish nine bilateral transport associations in order to send the published message for all lights to turn on. Communicating devices must be configured with each other's IP address and enrolled identity so, for n devices, both the configuration burden and traffic scale as O(n²). For example, when an "all" event is triggered, every light's radio will receive nine messages but discard the eight determined to be "not mine." If a device sleeps, is out-of-range, or has partial connectivity, additional application-level mechanisms have to be implemented to accommodate it.¶

MQTT and other broker-based pub/sub approaches mitigate this by adding a broker (Figure 4). Each entity makes a single TCP transport connection with the broker and tells the broker to subscribe it to topics. Then the kitchen switch uses its single transport session to publish commands to topic kitchen/counter, topic kitchen or all. The kitchen counter light uses its broker session to subscribe to those same three topics. The kitchen ceiling lights subscribe to topics kitchen ceiling, kitchen and all while den ceiling lights subscribe to topics den ceiling, den and all. Use of a broker reduces the configuration burden from O(n²) to O(n): 18 transport sessions to 11 for this simple example but for realistic deployments the reduction is often greater.¶

There are other advantages: besides their own IP addresses and identities, devices only need to be configured with those of the broker. Further, the broker can store messages for temporarily unavailable devices and use the transport session to confirm the reception of messages. This approach is popular because the pub/sub application layer protocol provides an easy-to-use API and the broker reduces configuration burden while maintaining secure, reliable delivery and providing short-term in-network storage of messages. Still the broker implementation doubles the per-device configuration burden by adding an entity that exists only to implement transport and traffic still scales as O(n²), e.g., any switch publishing to all lights results in ten (unicast) message transfers over the wifi network. Further, the broker introduces a single point of failure into a network that is richly connected physically.¶

Clearly, a transport protocol able to exploit a physical network's broadcast capabilities would better suit this problem. (Since unicast is just multicast restricted to peer sets of size 2, a multicast transport handles all unicast use cases but the converse is not true.)¶

More general solutions for this communications paradigm are possible by moving the view of the problem from message exchange to the concept of coordinating shared objectives. In the distributed systems literature, communication associated with coordinating shared objectives has long been modeled as distributed set reconciliation [WegmanC81][Demers87]. In this approach, each domain of discourse is a named set, e.g., myhouse.iot. Each event or action, e.g., a switch button press, is added as a new element to the instance of myhouse.iot at its point of origin then the reconciliation process ensures that every instance of myhouse.iot has this element. In 2000, [MINSKY03] developed a broadcast-capable set reconciliation algorithm whose communication cost equaled the set instance differences (which is optimal) but its polynomial computational cost impeded adoption. In 2011, [DIFF] used Invertible Bloom Lookup Tables (IBLTs) [IBLT][MPSR] to create a simple distributed set reconciliation algorithm providing optimal in both communication and computational cost. DeftT uses this algorithm (see Section 3.2) and takes advantage of IPv6's self-configuring link local multicast to avoid all manual configuration and external dependencies. This restores the system design to Figure 3 where each device has a single, auto-configured transport that makes use of the broadcast radio medium without need for a broker or multiple transport associations. Each button push is broadcast exactly once to be added to the distributed set. (See Section 3.5 to see how members fill in missing elements.)¶

1.3. Securing information

When pub/sub is implemented using conventional session-based transports multiple publications with independent topics and purposes are combined under a single session key, providing privacy by encrypting the sessions between endpoints. Credentials of Internet endpoints (e.g., a website) are usually attested by a third party certificate authority (CA) and bound to a DNS name. Endpoint credentials may also be bound to a device by using BRSKI [RFC8995] to obtain the (local) domain's root certificate then using a TLS connection to obtain the end-entity (EE) certificate. Each secure transport association requires the exchange of these credentials which allows for secure exchange of a nonce symmetric key. For example, in Figure 4 each transport session is a separate security association where each device validates the broker's credential and the broker has to validate each device's. Secured transport associations are between two enrolled devices (protecting against outsider and some MITM attacks) but, once the transport session has been established there are no constraints whatsoever on what devices can say. Clearly, this does not protect against the insider attacks that currently plague OT, e.g., [CHPT] description of a lightbulb taking over a network. For example, the basic function of a light switch requires that it be allowed to tell a light to turn on or off but it almost certainly shouldn't be allowed to tell the light to overwrite its firmware (fwupd), even though "on/off" and "fwupd" are both standard capabilities of most smart light APIs. Once a TLS session is established, the transport handles "fwupd" publications the same way as "on/off" publications. Such attacks can be prevented using trust management that operates per-publication, using rules that enable the "fwupd" from the light switch to be rejected. Combining per-publication trust decisions with many-to-many communications over broadcast infrastructure requires per-publication signing rather than session-based signing. More on how DeftT addresses past and current threats is in Section 1.6 and Section 7.¶

Securing each publication rather than the session it arrives on deals with a wider spectrum of threats while avoiding the quadratic session state and traffic burden. In OT, valid messages conform to rigid standards on syntax and semantics [IEC61850][ISO9506MMS][ONE][MATR][OSCAL][NMUD][ST][ZCL] that can be combined with site-specific requirements on identities and capabilities to create a system's communication rules. These rules can be employed to secure publications in a trust management system such as [DLOG] where each publisher is responsible for supplying all of the "who/what/where/when" information needed for each subscriber to prove the publication complies with system policies.¶

Instead of vulnerable third-party CAs [W509], domains employ a local root of trust and locally created certificates. Communication rules are expressed in a declarative language that can be validated for consistency and completeness then converted to a compact runtime form which is authorized and secured via signing with the system trust anchor. This communication schema is distributed as a certificate that can be validated using on-device trusted enclaves [TPM][HSE][ATZ] as part of the device enrollment process. See Section 5 and Section 7.¶

Figure 5 shows a simplified version of a communication schema for Figure 3. These member identities are simple two-level signing chains, the domain trust anchor signs the identities (device certs). The notation "<=" should be read as "is signed by" and the "& {}" notation indicates particular constraints. We give switches and lights different attributes in their identities ("switch" or "light") and then require that publications that send commands to turn on or off must be signed by switch identities. We limit light identities to signing publications that indicate their on or off status. When a light constructs a publication, the "room" and "loc" fields will be set to its identity's "_myroom" and "_myloc" attributes. When validating a light publication, "room" and "loc" must equal the pub signer's "_myroom" and "_myloc" and the "arg" must equal either "on" or "off". These rules specify that all certificates have three final components, the first of which is always "KEY"; the other two follow requirements within the trust engine (DeftT's approach is detailed in Section 3.3).¶

_domain: "myLights"              // trust domain name
domainCert: _domain/_certSuffix  // domain trust anchor

// publication format
#lsPub: _domain/room/loc/arg/_ts & {_ts: timestamp()}

// what switches and lights can say
rooms:  "kitchen"|"den"
switch: #lsPub & {room: rooms|"all", arg: "turnOn"|"turnOff"} <= switchCert
light:  #lsPub & {room: _myroom, loc: _myloc, arg: "on"|"off"} <= lightCert

// device certificate formats
deviceCert: _domain/_type/_myroom/_myloc/_certSuffix <= domainCert
switchCert: deviceCert & {_type: "switch"}
lightCert:  deviceCert & {_type: "light"}

// fields added to all cert names by cert builder
_certSuffix: "KEY"/_/_

The administrator of the "myLights" domain compiles the text version of the schema into binary and makes a schema cert, then makes a trust anchor, kitchen and den switch identities, four kitchen ceiling light identities, one kitchen counter light identity, and four den ceiling light identities. (Utilities for this are provided at [DCT].) The trust anchor private key is used to sign all of these and each device is preconfigured with the trust anchor cert, the schema cert, and its device cert (along with the device cert private key).¶

In DeftT's publication-based transport, the schema is used to both construct and validate publications, guaranteeing that all parts of the system always conform to and enforce the same rules, even as those rules evolve to meet new threats (more in Section 4.1). DeftT embeds the trust management mechanism described above directly in the publish and subscribe data paths as shown below:¶

This approach extends LangSec's [LANGSEC] "be definite in what you accept" principle by using the authenticated common rules of the schema for belt-and-suspenders enforcement at both publication and subscription functions of the transport. If an application asks the Publication Builder to publish something and the schema shows it lacks credentials, an error is thrown and nothing is published. Independently, the Publication Validator ignores publications that:¶

• don't have a locally validated, complete signing chain for the credential that signed it¶
• the schema shows its signing chain isn't appropriate for this publication¶
• have a publication signature that doesn't validate¶

Since an application's subscriptions determine which publications it wants, only certificates from chains that can sign publications matching the subscriptions need to be validated or retained. Thus a device's communication state burden and computation costs are a function of how many different things are allowed to talk to it but not how many things it talks to or the total number of devices in the system. In particular, event driven, publish-only devices like sensors will spend no time or space on validation. Unlike most 'secure' systems, adding additional constraints to schemas to reduce attack surface results in devices doing less work.¶

1.4. Defined-trust Communications Domains

A Defined-trust Communications Limited Domain (or simply, trust domain) is a Limited Domain where communications are via DeftT (Figure 7) and all members are configured with the same trust anchor and schema (or some subset of the domain schema) as well as an individual schema-conformant DeftT identity cert chain that terminates at the trust anchor and the private key corresponding to the identity chain's leaf cert. The particular rules for any deployment are application-specific (e.g., Is it home IoT or a nuclear power plant?) and site-specific (specific form of credential and idiosyncrasies in rules) which DeftT accommodates by being invoked with a ruleset (schema) particular to a deployment. We anticipate that efforts to create common data models (e.g., [ONE]) for specific sectors will lead to easier and more forms-based configuration of DeftT deployments.¶

A trust domain is perimeterless and may operate over one or more subnets, sharing physical media with non-members. Domain members use DeftT to publish and subscribe using Publication Builders and Validators as shown in Figure 6. Publications become the elements of a set, or named collection, that is synchronized across each subnet of a trust domain that is using the same schema or schema subset. DeftT uses a distributed set reconciliation protocol on each collection and each subnet independently. A DeftT's set reconciliation protocol operates in a sync zone which is defined on a single subnet where members use the same communication schema or communication schema subset. Every DeftT maintains at least two collections: msgs for Publications constructed from application messages and cert where member signing chains are published.¶

Trust domains are extended across physically separated subnets, subnets using different media and/or groups of members using distinct subsets of the domain schema by relays. (see Section 3.7) Relays have a DeftT in each sync zone and pass Publications between them as long as the Publications are valid at both the receiving and sending DeftTs. Since set reconciliation does not accept duplicates, relays are powerful elements in creating efficient configuration-free meshes. In Figure 8, different sync zones are denoted by shape and could represent different colocated media (e.g. bluetooth, wifi, ethernet) or physically distant sync zones. The shaded triangles represent DeftTs in a sync zone with only relay members, as for a unicast link between two relays. The set reconciliation protocol ensures that items only transit a sync zone once; an item must be specifically requested in order to be transmitted.¶

Any part of a verifiable defined-trust identity can be used in the delineation of sync zones, e.g. specific component(s) of identity names allowed to sign Publications can be constrained to be identical so that Publications are effectively only relayed to a particular "group" as identified by those components. This is enforced via the secured schema, i.e., non-"group" Publications will not validate. Relay identities have a relay capability (an attribute with a special meaning within DeftT) in their signing chain so that they can use specialized certificate and Publication movement while being prohibted from obtaining Publication encryption/decryption keys. There are four subnets in Figure 8, represented as different shapes. The subnets could be using different media, different group addresses, or be physically remote. A relay has a DeftT in two or more subnets of the Domain, e.g. one of these covers subnets 1, 3, and 4. Conceptually, relay DeftTs maintain a collection in their subnet, e.g. subnet1, on behalf of the members in its other subnets, e.g., subnets 3 and 4. Relay internals copy Publications to the other DeftTs but only after validating them using the (sub)schema in use on the destination DeftT. Relay discussion is in Section 3.7 and Section 6.¶

1.5. Discovery

Conventional IP services are discovered via well-known IP ports that are registered with IANA [RFC6335]. DeftT is not a service, so it doesn't require a port allocation. Instead, DeftT does peer discovery using shared security information. In particular, each DeftT is a member of a trust domain and communicates within a single sync zone where all members use the same schema cert, which is signed by the common trust anchor (see Section 1.4). The SHA256 thumbprint Section 2 of the schema cert for a sync zone is a compact, unique identifier.¶

 8|  4  |  4  |  112   |
--+-----+-----+--------+
FF|flags|scope|group ID|
--+-----+-----+--------+

FF|01|02|<lower 112 bits of schema cert thumbprint>

1.6. Design rationale

DeftT's Publication, PDU and serialization formats were strongly influenced by the LangSec observation that most security issues are due to improper input handling. For example, [LangSecErr] (section II) found that this class of errors accounted for 75% of the 47 OpenSSL security vulnerabilities reported in the 18 months following 2015-1-1. Also, as of 2023-7-5, all 25 of the protobuf CVEs listed in the NIST National Vulnerability Database are of this class.¶

[LangSecErr] suggests these vulnerabilities could have been avoided by designing the protocol following three rules:¶

The acceptable input to a program should be:¶

1. well-defined (i.e., via a grammar)¶
2. as simple as possible (on the Chomsky scale of syntactic complexity)¶
3. fully validated before use (no "shotgun parsing")¶

1.7. Current status

An open-source Defined-trust Communications Toolkit [DCT] with an example implementation of DeftT is maintained by the corresponding author's company. [DCT] currently has examples of using DeftT to implement secure brokerless message-based pub/sub using multicast UDP/IPv6 and unicast {UDP/TCP}/IPv6 and includes extending a trust domain via a unicast connection or between two broadcast network segments.¶

Massive build out of the renewable energy sector is driving connectivity needs for both monitoring and control. Author King's company, Operant, is currently developing extensions of DeftT in a mix of open-source and proprietary software tailored for commercial deployment in support of distributed energy resources (DER). Current small scale use cases have performed well and expanded usage is underway. Pollere is also working on home IoT uses. The development philosophy for DeftT is to start from solving useful problems with a well-defined scope and extend from there. As the needs of our use cases expand, the Defined-trust communications framework will evolve with increased efficiencies. DeftT's code is open source, as befits any communications protocol, but even more critical for one attempting to offer security. DCT itself makes use of the open source cryptographic library libsodium [SOD] and the project is open to feedback on potential security issues as well as hearing from potential collaborators.¶

3.1. Inside DeftT

DeftT's example implementation [DCT] is organized in functional library modules that interact to prepare application-level information for transport and to extract application-level information from packets, see Figure 10. Extensions and alternate module implementations are possible but the functionality and interfaces must be preserved. Internals of DeftT are completely transparent to an application and the example implementation is efficient in both lines of code and performance (see Section 3.8). The schema determines which modules are used (Figure 11). A DeftT participates in two required collections and may participate in others if required by the schema-designated cryptographic methods (see #distributors).. One of the required collections, msgs, contains Publications constructed from application messages. The other required collection, cert, contains the certificates of the trust domain. Specific cryptographic methods may require group key distribution in descriptively named collection keys.¶

A shim serves as the translator between application semantics and the named information objects (Publications) whose format is defined by the schema. The syncps module is the set reconciliation protocol used by DeftT (see Section 3.2). New signature managers, distributors, and face modules may be added to the library to extend features. More detail on each module can be found at [DCT] in both code files and documents.¶

The signing/validation modules (signature managers) are used for both Publications and cAdds. Figure 11 shows validator specifications for the two required Publication collections, one for Publications carrying application messages and one for Publications carrying certificates. The "#pduValidator" specifies AEAD, which requires a symmetric key, an example of a sigmgr that requires another distributor module (see #(group-key-distributors)) that is automatically instantiated at run-time. Changing a pduValidator from "EdDSA" to "AEAD" in a domain's communication schema is all that needs to be done to change from signed, cleartext PDUs to PDUs encrypted with a periodically regenerated shared key where only the minimal information of trust domain identifier and collection type is exposed.¶

#msgsValidator: "EdDSA"
#certValidator: "EdDSA"
#pduValidator:  "AEAD"  //for cAdds

Following good security practice, DeftT's Publications are constructed and signed early in their creation, then are validated (or discarded) early in the reception process.The schemaLib module provides certificate store access throughout DeftT along with access to distributors of group keys, Publication building and structural validation, and other functions of the trust management engine.¶

3.2. syncps: a set reconciliation protocol

DeftT requires a method or protocol that keeps collections of Publications synchronized. Required functionality for such a protocol can be understood through the example of the syncps protocol included in the example implementation. The syncps protocol uses IBLTs to solve the multi-party set-difference problem efficiently without the use of prior context and with communication proportional to the size of the difference between the sets being compared. The state of a local collection is encoded in an IBLT. IBLTs have been used for set reconciliation in blockchain systems, specifically in Bitcoin Cash, contributing optimizations available as open-source code [Graphene19] that syncps can utilize.¶

A syncps announces its local collection state (set of currently known Publications) by sending a cState (Section 3.3.1.1) that also serves as a query for additional data not reflected in its local state. Receipt of a cState performs three simultaneous functions: (1) announces new Publications, (2) notifies of Publications that member(s) are missing and (3) acknowledges Publication receipt. The first may prompt the recipient to share its cState to get the new Publication(s). The second results in the recipient sending a cAdd (Section 3.3.1.2) containing all the locally available missing Publications that fit. The third is used optionally and may result in a progress notification sent to other local modules so anything waiting for delivery confirmation can proceed.¶

On broadcast media, syncps uses any cStates it hears to reduce (suppress) sending excess cStates and listens for cAdds that may add to its collection. This means that one-to-many Publications cause sending a single cState and a single cAdd independently of the number of members desiring the Publication (the theoretical minimum possible for reliable delivery). The digest size of a cState can be controlled by Publication lifetime, dynamically constructing the digest to maximize communication progress and, if necessary for a large network, dynamically adapting topic specificity.¶

A cAdd with new Publication(s) responds to a particular cState as per (Section 3.3.1.2 item 1). Any DeftT that is missing a Publication (due to being out-of-range, asleep, channel errors, etc.) can receive it from any other DeftT. A syncps will continue to send cAdds as long as cStates are received that are missing any of its active Publications. This results in reliability that is subscriber-oriented, not publisher-oriented, kept efficient with protocol features that prevent multiple redundant broadcasts. The example implementation of syncps prevents redundant broadcasts by having originating publishers send their responding Publications immediately while others delay before supplying missing Publications, canceling if a responding cAdd is overheard. Other approaches are possible.¶

The collection synchronization work of a syncps module is shown as a state diagram in Figure 12. When a new syncps is started, it always sends its local cState (starts unsuppressed) on the network and sets an expiration timer for the cState. If this timer expires, the "new local cState" actions are repeated and the cState may be suppressed (thus not sent). For most collections, the initial cState will show an empty collection (certificate collections will have the local identity chain). The events that can move the collection forward are (1) the arrival of a cState from the network, (2) the arrival of a cAdd from the network whose csID matches a hash value stored from a previously received or sent cState, or (3) arrival of a new Publication from its shim. For an arriving cAdd, each Publication is extracted, validated, and passed to any registered subscriber callback(s). Non-validating packets are silently discarded (may optionally set alerts or count discards). Reception of new Pub(s) may cause an application process to create and add new Publications to the local collection. Sending of new Pubs is deferred until the entire cAdd has been processed. If there are no new Pubs to send, syncps moves to its "set sendCStateTimer" state where a cancelable sendCState timer is set to the estimated dispersion delay of this local subnet. (Dispersion should be << cState lifetime. More on dispersion delay in Section 3.5.)¶

Since new Publications are always eligible to send, if any were created while in "process cAdd", the next state is "process Pubs to send" with csID set to the csID field of the cAdd on entry to "process cAdd". In "process Pubs to send" any pending sendCState will be canceled and eligible Pubs are packaged as content for a new cAdd. Packaged Publications are subject to a hold time (during which they are ineligible to send) of twice the dispersion delay to avoid responding to cStates sent before reception of a cAdd containing the Pub. If there are Pubs to send, a cAdd with that content and the passed in csID is sent, then the set sendCStateTimer state is entered; when there are no Pubs to send, the action moves there directly.¶

Another path to exit the "wait for event" state is reception of a cState which moves to "process cState" where incoming cStates are recorded, If this cState matches the local cState, the syncps returns to the wait state. Otherwise, the IBLT is extracted from the cState, an IBLT is computed on the local Pub collection, and they are peeled to find the ones the received cState has that are not in the local collection ("needs") and the ones that are in the local collection and not in the received cState ("haves"). Syncps enters the "process Pubs to send" state with csID set to the hash of the cState's name. Eligible Pubs are "haves" that do not have a hold time set and locally generated Publications are sent preferentially. Publications obtained from others are not immediately eligible to send; members delay to give the originator time to respond, sending these when further cStates indicate a member continues to need them.¶

A syncps also exits the wait state when the attached shim has a Publication to send. Since a new Pub will not appear in any previously issued cState, any one can be used, including one issued locally. In "get stored cState", the best one (the most recent cState from the network if available) is retrieved and passed to the "peel IBLT values" state where its cState.csID is used for the csID value. There will always be at least the one new Publication to send in this case.¶

This state diagram is intended to capture the major functionality of a syncps module while excluding excessive detail. In particular, Figure 12 does not show "housekeeping" tasks on the collection, e.g., removal of expired Publications.¶

3.3. DeftT formats

All DeftT Information is represented using TLV (Type, Length, Value) tuples. Types can either be containers (they contain a concatenated sequence of TLVs) or leaves (they contain a single non-TLV value with well-defined semantics and serialization). All TLVs have a boolean 'valid()' method that returns 'true' if and only if their content satisfies all the constraints associated with the TLV's type. For container types this means, at minimum, that the sum of all the enclosed TLV Lengths and header sizes exactly equals the Length of the container and that the valid() method of each of the enclosed TLVs returns true. Most container types have additional constraints on the type, ordering and value of the enclosed TLVs that are described below.¶

5 (cState) size 128:
| 7 (Name) size 116:
| | 8 (Generic) size 8:  55d5 7f99 7d8d ba91
| | 8 (Generic) size 4:  cert
| | 8 (Generic) size 98:  8201 dd76 eb0f 46ed  89a8 8101 dd76 eb0f  46..
| |                       beb5 9922 fdd6 7401  cbbe 5bc1 5b57 1c63  84..
| |                       79aa ca17 8501 cbbe  5bc1 5b57 1c63 8801  ce..
| |                       92f8
| 10 (Nonce) size 4:  8b9f 8134
| 12 (Lifetime) size 2:  4789

6 (Data) size 561:
| 7 (Name) size 22:
| | 8 (Generic) size 8:  55d5 7f99 7d8d ba91
| | 8 (Generic) size 4:  cert
| | 35 (csID) size 4:  f6d7 3d84
| 20 (MetaInfo) size 3:
| | 24 (ContentType) size 1:  42 (CAdd)
| 21 (Content) size 489:
         ... (489 bytes of Content elided)
| 22 (SigInfo) size 3:
| | 27 (SigType) size 1:  9 (RFC7693)
| 23 (SigValue) size 32:  af8e 1412 e659 103f  5237 f1e1 0e7b 0af8  9c..

6 (Data) size 216:
| 7 (Name) size 68:
| | 8 (Generic) size 4:  iot1
| | 8 (Generic) size 4:  lock
| | 8 (Generic) size 7:  command
| | 8 (Generic) size 3:  all
| | 8 (Generic) size 4:  lock
| | 8 (Generic) size 17:  p38863@aphone.local
| | 37 (SequenceNum) size 4:  b4a1 ea2a
| | 37 (SequenceNum) size 0:
| | 36 (Timestamp) size 7:  23-09-18@19:40:45.591793
| 20 (MetaInfo) size 3:
| | 24 (ContentType) size 1:  0 (Blob)
| 21 (Content) size 32:  Msg #3 from operator:alice-38863
| 22 (SigInfo) size 39:
| | 27 (SigType) size 1:  8 (EdDSA)
| | 28 (KeyLocator) size 34:
| | | 29 (KeyDigest) size 32:  7096 5de9 6848 7543  d2c8 e459 24fb 7b0..
| 23 (SigValue) size 64:  61b3 fc3c 03df 2c89  7a0c ddae 27a2 f883  dd..
|                         2699 899f 1c91 46c1  3127 9da8 8948 e783  68..

6 (Data) size 214:
| 7 (Name) size 69:
| | 8 (Generic) size 4:  iot1
| | 8 (Generic) size 4:  lock
| | 8 (Generic) size 5:  event
| | 8 (Generic) size 4:  gate
| | 8 (Generic) size 6:  locked
| | 8 (Generic) size 17:  p59280@rpi2.local
| | 37 (SequenceNum) size 4:  e131 5a4b
| | 37 (SequenceNum) size 0:
| | 36 (Timestamp) size 7:  23-09-18@19:40:45.594867
| 20 (MetaInfo) size 3:
| | 24 (ContentType) size 1:  0 (Blob)
| 21 (Content) size 29:  Msg #3 from device:gate-59280
| 22 (SigInfo) size 39:
| | 27 (SigType) size 1:  8 (EdDSA)
| | 28 (KeyLocator) size 34:
| | | 29 (KeyDigest) size 32:  3dde 0f21 beae 2c20  3ea3 5c2e 77ca 9d4..
| 23 (SigValue) size 64:  3913 011d 7e74 807c  94b5 e725 a8e7 5b2f  09..
|                         bc99 9c8b fa9f f929  4722 f23a 1fbe cd84  b6..

6 (Data) size 240:
| 7 (Name) size 50:
| | 8 (Generic) size 4:  iot2
| | 8 (Generic) size 6:  device
| | 8 (Generic) size 9:  frontdoor
| | 8 (Generic) size 3:  KEY
| | 8 (Generic) size 4:  0eaf f793
| | 8 (Generic) size 3:  dct
| | 36 (Timestamp) size 7:  23-02-18@18:17:46.088971
| 20 (MetaInfo) size 3:
| | 24 (ContentType) size 1:  2 (Key)
| 21 (Content) size 32:  de19 4605 7f77 a7bd  1317 de41 002c fe15  1bc..
| 22 (SigInfo) size 81:
| | 27 (SigType) size 1:  8 (EdDSA)
| | 28 (KeyLocator) size 34:
| | | 29 (KeyDigest) size 32:  8c7f 1de9 ebc9 17b6  a8e9 dce9 056a 74c..
| | 253 (Validity) size 38:
| | | 254 (NotBefore) size 15:  20230219T021746
| | | 255 (NotAfter) size 15:  20240219T021746
| 23 (SigValue) size 64:  c8b9 5883 4b9a 8aac  9ad0 e5e4 5eef 0a18  4b..
|                         1b3a 1574 58d4 0528  1740 883e d90c 836f  ed..

3.4. Application and network interface

Figure 9 and Figure 10 show the blocks and modules application information passes through in DeftT and may be useful references for this subsection. In a trust domain, applications pass information to be communicated to their DeftT, which packages it into Publication(s) that are added to the local collection copy. These Publications are also sent in a PDU via the system network interface to be received by other members of the domain which add the Publications to their local collections. If a received Publication matches a subscription, the information it contains is passed to the application. (For more detail, see the library at [DCT].) DeftT is organized into modules that perform its tasks. A DeftT shim exchanges information with applications. The example implementation [DCT] provides a message-based publish/subscribe (mbps) API that exchanges messages with the application and Publications with the sync protocol. (Other APIs are possible.) DeftT startup begins when a shim object is instantiated by the application and given its identity bundle. Startup includes creating a certificate distributor and, optionally, group key distributors, depending on schema-specified signing. After startup, the msgs syncps of each member will maintain a cState containing the IBLT of its view of the collection. (In the stable, synchronized state, all members of a collection will have the same IBLT.)¶

Applications subscribe to all messages, or to a subset by topic, by passing a callback function to the mbps subscribe method. Application subscriptions are turned into syncps subscriptions via mbps. An application with new information to communicate passes the formatted topic items as parameters and the other content in a message via a publish call. Only the topic components and the message, if any, are passed between the application and mbps. Mbps adds mbps-specific components to the parameter list and invokes a schemaLib method that builds a valid (according to the schema) Publication and can be passed to syncps to publish. Messages that exceed the content size limits of a single Publication are segmented by mbps and carried in multiple Publications. If a member's identity chain lacks the attributes required for a specific Publication, no Publication is built. The Publication is signed using the sign method of the appropriate sigmgr and passed to syncps.¶

syncps adds this Publication to its collection and updates its IBLT to contain the new Publication. Since its application just created it, the Publication is a new addition and thus is always a response to the current cState. The Publication is packaged into a cAdd and signed using the sign method of the designated sigmgr and passed to the face. The updated IBLT is packaged into a new cState that is handed to the face.¶

Trust domain members only process cAdds that share their trust domain identifier (Section 3.3.1.1 and Section 3.3.1.2). When a new cAdd is received at a member, the face ensures it matches an outstanding cState and, if so, passes it on to its matching syncps(es). Syncps validates (both structurally and cryptographically) the cAdd using the appropriate sigmgr's validate and continues, removing Publications, if valid. Each Publication is structurally validated via a sigmgr and valid Publications are added to the local collection and IBLT. syncps passes this updated cState to the local face. If this Publication matches a subscription it is passed to mbps, invoking the sigmgr's decrypt if the Publication is encrypted. (Publication decryption is not available at relays.) mbps receives the Publication and passes any topic components of interest to the application along with the content (if any) via the callback registered when it subscribed. (If the original content was spread across Publications, mbps will wait until all of the content is received. The sCnt component of a mbps Publication Name is used for this.)¶

3.5. Synchronizing a collection

DeftT works on unicast (as a special case of multicast) links, but is designed to take full advantage of a multicast subnet (e.g., link-level IPv6 multicast on broadcast media) with syncps orchestrating collection-based communications. Members' syncps modules interact on a multicast subnet to keep their collections synchronized. The following example illustrates member actions and communications to synchronize a collection (see also the sequence diagram in Figure 13).¶

Starting with all members connected to the collection (having confirmed publication of their identity credentials) and with an empty msgs collection (i.e., no applications have active Publications), member2's application passes content to its DeftT via an mbps.publish(). The content is packaged into a Publication (p1) and passed to syncps which creates and sends a cAdd PDU. The cAdd uses a hash of the shared (empty) cState as its cState identifier (third component of the Name -- see Section 3.3.1.2 item 1) to indicate the Publication(s) it carries are additions to the collection in that state. Member2's new local cState (containing p1) is scheduled to be sent at a delay of the subnet's dispersion time (d) plus a small random value (r). Dispersion time is an estimate of the expected time for a cAdd to reach every member's collection. It may be a fixed or adaptive estimate and syncps is robust to inaccuracies: an overestimate may lead to longer delays and an underestimate may lead to more cState traffic on the channel. Members receive and validate the cAdd, then extract and validate p1, passing it to subscriptions. Each member schedules a sendCState after a small random delay r. (Scheduling a new sendCState cancels any pending sendCState.) When the sendCState timer expires, a new local cState is created with the IBLT of the collection (which will contain p1). This cState's expiration time is scheduled (value significantly longer than d) and the member sends the cState unless it is suppressed. DeftT suppresses cStates that are identical to one that has already been heard twice. If member2 is waiting to confirm p1, it can do so with the first of these cStates it receives. In Figure 13, member6 did not receive the cAdd but reception of one of the new cStates shows the presence of p1 so member6 immediately sends its own local cState (which has an empty collection, lacking member2's Publication). In this example, all members receive member6's cState, but member2, as p1's originator, responds preferentially and sends p1 in a new cAdd immediately. All other members set a timer (to d+r) to send p1. That timer is cancelled if the member receives a cAdd responding to member6's cState that contains p1. Meanwhile, member6 receives the new cAdd, adds p1 to its collection and schedules a new cState for delay r. That cState will be suppressed as it matches those already sent by the other members. Now the distributed collection is synchronized with a state of one Publication (p1). If no other application content is created, cStates will be sent at ~cStateLifetime. On the channel, we will see one cState per ~cStateLifetime since each overlaps enough to suppress others. When p1 expires, it will be removed at each local collection and the subsequent cState will show an empty collection.¶

Although Figure 13 shows one Publication at a time for clarity, the logic applies if multiple members are publishing simultaneously or at close intervals (less than d or the cStateLifeTime). Distributed collections are always moving toward synchronization but during periods of intense interaction, times when all members are synchronized may be infrequent; this is not problematic.¶

3.6. Distributors

Distributors implement services a Deft requires for its operation. Distributors optional to general operation are specified in the communication schema.¶

3.7. Schema-based information movement

Although the Internet's transport and routing protocols emphasize universal reachability with packet forwarding based on destination, a significant number of applications neither need nor desire to transit the Internet (e.g., see [RFC8799]). This is true for a wide class of OT applications. Further, liberal acceptance of packets while depending on the good sending practices of others leaves critical applications open to misconfiguration and attacks. Internet protocols use header information to tell them how to forward packets; A DeftT PDU's header only contains a sync zone id and a collection name. Each DeftT has a trust management engine with a copy of rules (a schema or subschema). DeftT only creates and moves its Publications in accordance with the fully specified communication schemas and never moves a PDU between sync zones. This approach differs in both intent and execution from Internet forwarding. It may not be appropriate for all use cases but offers new opportunities to address the specific security requirements of many Limited Domain use cases.¶

DeftT PDUs on the same subnet may be in different sync zones or trust domains and DeftT sync zones in the same trust domain may be on different subnets. In some cases, it is useful to define sync zones whose DeftTs have a compatible, but more limited, version of the trust domain's communication schema which is itself complete as a communication schema. This subschema concept was introduced in Section 1.3 and further discussed in Section 4. "Compatible" means there is at least one Publication type and associated signer specification in common or one schema may be a subset of the other. Different subschemas may be deployed for sync zones on the same subnet or on different subnets. A subschema cert will have a different thumbprint from that of the full trust domain and different sync zones can be identified by the thumbprint of the (sub)schema in use.¶

For example, a unicast link may be used to connect two remotely located subnets of the same trust domain and only certain types of Publications should pass through the unicast link. Relays can be used where the DeftTs on the unicast link have a restricted subschema (e.g. Figure 14-right). Further, different sync zones on the same subnet might be used where certain members have more limited access, either due to the technology of their devices or to restrict their access (e.g., guests of a network). Relays could limit Publications simply by filtering Publications or subscribing to subsets of Publications but use of a subschema in different sync zones provides enforcement of Publication movement.¶

Both cStates and cAdds contain their sync zone id and are not moved between subnets while Publications are defined in the trust domain's communication schema and can move to any DeftT that can validate them. In the case of DeftTs on the same subnet but in with different (sub)schema certs, the cState and cAdd PDUs are differentiated by the sync zone id (thumbprint of the (sub)schema certificate as in Section 3.3.1.1 item C1). The sync zone id is used at the face module to determine whether or not to process a PDU. A DeftT's syncps manages a particular collection on a single subnet. Relays move Publications between separate sync zones of the same trust domain by moving Publications between the relay's multiple DeftT instances.¶

A relay is implemented [DCT] as an application running on a device with a DeftT interface in each sync zone (two or more) Figure 14. Each DeftT participates using a communication identity valid for the schema used by the DeftT. Only Publications (including certs) are relayed between DeftTs and the Publication must validate (to the extent possible) against the schema of each DeftT. Consequently cAdd encryption is unique per sync zone while Publication encryption holds across the domain.¶

Since relay applications merely pass Publications in the msgs collection, their DeftT API module (a "shim", see Section 3.1) performs pass-through of valid Publications. As a consequence and a further security measure for boundary devices, relays have no need for Publication encryption keys; this is enforced by use of a capability cert in relay identity chains. (The group symmetric key is never given to an identity with the relay capability in its chain.) For example, if we added a relay definition to the example of Figure 5:¶

rlyCert: _domain/"relay"/_myid/_certFormat <= rlyCap
capCert: _domain/"CAP"/capId/capArg/_certFormat
rlyCap:  capCert & {capId: "RLY", capArg: _} <= domainCert

The relay of Figure 14-left is on three separate wireless subnets. If all three DeftTs are using an identical schema, a new validated cert added to the cert store of an incoming DeftT is then passed to the other two, which each validate the cert before adding to their own cert stores (superfluous in this case, but not a lot of overhead for additional security). When a valid Publication is received at one DeftT, it is passed to the other two DeftTs to validate against their schemas and published if it passes.¶

A relay may have different identities and schemas for each DeftT, but its DeftTs must have the same trust anchor and schemas that are identical copies, proper subsets or overlapping subsets of the domain schema. Publications that are undefined for a particular DeftT are silently discarded if they do not validate upon relay, just as they are when received from a face. This means the relay application of Figure 14-left can remain the same but Publications will only be published to a different subnet if its DeftT has that specification in its schema. In addition, relays may filter Publications at the application level or restrict subscriptions on some of their DeftT interfaces. Figure 14-right shows extending a trust domain geographically by using a unicast connection (e.g., over a cell line or tunnel over the Internet) between two relays which also interface to local broadcast subnets. Everything on each local subnet shows up on the other. A communication schema subset could be used here to limit the types of Publications sent on the remote link, e.g., logs or alerts. Using this approach in Figure 14-right, local communication for subnet 1 can be kept local while subnet 2 might send commands and/or collect log files from subnet 1.¶

More generally, relays can form a mesh of broadcast subnets with no additional configuration (i.e., relays on a broadcast network do not need to be configured with others' identities and can join at any time). The mesh is efficient: Publications are only added to an individual DeftT's collection once regardless of how it is received. Relays with overlapping broadcast physical media will only add a Publication to any of its DeftTs once; syncps ensures there are no duplicates. More on the applicability of DeftT meshes is in Section 6.¶

3.8. Performance

Measurements and profiling have been performed using examples from the open source [DCT] proof-of-concept codebase (v11.2) running on an Apple M1-max and an x86 linux-based machine. On both machines the code was compiled with clang-16 at optimization level -O3 but no additional or platform-dependent flags. An application examples/hmIoT/app2.cpp [DCT] with an operator role issue commands to app2.cpp's with device roles which are expected to perform the operation (e.g., lock the door) and report their subsequent status (e.g., locked). The schema used calls for EdDSA signing of Publications and AEAD encryption of cAdds. Measuring the time from the appearance of the cAdd carrying the command publication until the appearance of the cAdd carrying the status publication captures the time to receive a cAdd, process its Publication, convert the resulting application message into a Publication and then add the Publication to its local collection and package it and send it in a cAdd. On ethernet with negligible transmission delay, the Apple platform takes ~200us and the linux x86 ~300us.¶

Memory use is dominated by the certificate store (the cert collection) and the msgs collection of Publications carrying application content. In the former case, each certificate is ~300 bytes and a member only stores its own signing chain as well as those of members that create Publications the communication schema says they can accept. Measurement has shown memory use to be consistently less than 2 Mbytes.¶

The home IoT example was profiled on an M1-max using Apple's instruments tool. The time spent was classified both by the DeftT action it was invoked in and the codebase it was part of (DeftT, the libsodium cryptographic library, or the Boost async I/O library interface the to system's UDP/IP networking and scheduling). For each action, times were normalized by the total time spent on the action. The numbers in the table below are percentages of the total time taken for the action in the first column. For each codebase, the number given is the percentage time taken by its busiest function. The name of the action's largest item, irrespective of codebase, shown in the final column.¶

Note that none of the absolute times are large. For example, the EdDSA validation that accounts for 94% of incoming Pub-to-msg handling takes 22 microseconds and the corresponding signature that's 68% of make msg-to-Pub takes 16 microseconds. In general, send and receive syscall handling dominates and these delays are experienced by any transport protocol. Roughly half of this appears to be due to Boost's lock-heavy "executor" model which might be fixed by switching to c++20 co-routines. A real-time OS like Zephyr would be expected to remove much of the remaining system cost.¶

Thanks to the availability of hardware AES accelaration on both Arm and Intel platforms, the time to perform the cAdd AEGIS crypto is quite small: 300ns to simultaneously validate and decrypt on either platform. These measurements convince us that users and industries promoting the use of signing in networking would do well to focus on accelerating asymmetric key functions.¶

Table 4 shows the fixed and per-byte signing and validation costs for the current reference implementation [DCT] as measured by its time_signing.cpp tool. Hardware accelerated AEGIS-128L and AEGIS-256, were added in Sep 13, 2023 libsodium 1.0.19 release. AEAD-AEGIS is 4.5x faster than AEAD-IETF on the Intel platform and 7.8x faster on the Arm platform. Given the rapid evolution of high quality crypto algorithms and implementations, it's important that infrastructure and application code that relies on crypto be able to keep up. It took less than a day to add an AEGIS sigmgr to [DCT] and a one line change to the home IoT schema to switch the home IoT apps to use it. Since the sigmgr API is generic, no app or DeftT code changed.¶

3.9. Congestion control

Each DeftT manages its collection on a single broadcast subnet (since unicast is a proper subset of multicast, a point-to-point connection is viewed as a trivial broadcast subnet), thus it only has to deal with that subnet's congestion. As described in the previous section, a device connected to two or more subnets may create DeftTs having the same collection name on each subnet with a Publication Relay between them but DeftT never forwards PDUs between subnets. It is, of course, possible to run DeftT over an extended broadcast network like a PIM multicast group but the result will generally require more configuration and be less reliable, efficient and secure than DeftT's self-configuring peer-to-peer Relay mesh.¶

DeftT sends at most one copy of any Publication over any fully connected subnet, independent of the number of publishers and subscribers on the subnet. Thus the total DeftT traffic on a subnet is strictly upper bounded by the application-level publication rate. As described in Section 3.2, DeftTs publish a cState specifying the set elements they currently hold. If a DeftT receives a cState specifying the same elements (Publications) it holds, it doesn't send its cState. Thus the upper bound on cState publication rate is the number of members on the subnet divided by the cState lifetime (typically seconds to minutes) but is typically one per cState lifetime due to the duplicate suppression. Each member can send at most one cAdd in response to a cState. This creates a strict request/response flow balance which upper bounds the cAdd traffic rate to (number of members - 1) times the cState publication rate. The flow balance ensures an instance can't send a new cState until it's previous one is either obsoleted by a cAdd or times out. Similarly a cAdd can only be sent in response to the cState which it obsoletes. Thus the number of outstanding PDUs per instance is at most one and DeftT cannot cause subnet congestion collapse.¶

If a Relay is used to extend a trust domain over a path whose bandwidth delay product is many times larger than typical subnet MTUs (1.5-9KB), the one-outstanding-PDU per member constraint can result in poor performance (1500 bytes per 100ms transcontinental RTT is only 120Kbps). DeftT can run over any lower layer transport and stream-oriented transports like TCP or QUIC allow for a 'virtual MTU' that can be set large enough for DeftT to relay at or above the average publication rate (the default is 64KB which can relay up to 5Mbps of Publications into a 100ms RTT). In this case there can be many lower layer packets in flight for each DeftT cAdd PDU but their congestion control is handled by TCP or QUIC.¶

4.1. Communication schemas

Defined-trust's use of communication schemas is influenced by [SNC][SDSI] and the field of trust management defined by Blaze et. al. [DTM] as the study of security policies, security credentials, and trust relationships. and refined by Li et. al. [DLOG]. Communications schemas also have roots in the trust schemas for Named-Data Networking, described in [STNDN] as "an overall trust model of an application, i.e., what is (are) legitimate key(s) for each data packet that the application produces or consumes." DeftT's use is based on an approach introduced in [NDNW] and extended with introduction of the declarative VerSec schema language in [DNMP][IOTK][DCT]. Here a schema is analogous to the plans for constructing a building. Construction plans serve multiple purposes:¶

1. Allow permitting authorities to check that the design meets applicable codes¶
2. Show construction workers what to build¶
3. Let building inspectors validate that as-permitted matches as-built¶

Construction plans get this flexibility from being declarative: they describe "what" to do, not "how" to do it. As noted on paragraph 4 of [DLOG]:¶

a declarative trust management specification based on a formal foundation guarantees all parties to a communication have the same notion of what constitutes compliance.¶

The schema is critical to Defined-trust Communications, adopting the more descriptive term communication schema (or just schema where its usage is clear) for the rules that define the communications of a trust domain. The schema details the meaning and relationship of individual components of the filename-like names (URI syntax [RFC3986]) of Publications and certificates. The intent for a domain's communications is expressed in fine-grained rules: "who can say what." Credentials that define "who" are specified along with complete definitions of "what". Defined-trust communications has been targeted at OT networking where administrative control is explicit and it is not unreasonable to assume that identities and communication rules can be securely configured at every device. A single schema, securely provided to all members, provides the same protection as dozens of manually configured, per-node ACL rules.¶

A VerSec compiler, schemaCompile, and examples are included with the Defined-trust Communications Toolkit [DCT]. The compiler checks the formal soundness of the VerSec text specification (case 1 above) then converts it to a signed, compact, binary form. The compiler also reports diagnostics and includes a digraph listing that can be used to confirm that the intent of the schema has been implemented (and to flag problems). The binary form is used by DeftT to build (case 2) or validate (case 3) the Publications (format covered in Section 3.3.1.3). As certificates are a type of Publication, they are distributed and validated using DeftT, though they are subject to additional constraints (Section 3.3.1.4).¶

4.2. VerSec schema description language

VerSec follows LangSec principles to minimize misconfiguration and attack surface. Its structure is amenable to a forms-based input or a translator from the structured data profiles often used by standards [ONE][RFC8520][ZCL]. As declarative languages are expressive and strongly typed, they can express the constructs of these standards in their rules. VerSec continues to evolve and add new features as its application domain is expanded; the latest released version is at [DCT]. Other languages and compilers are possible as long as they supply the features and output needed for DeftT.¶

VerSec describes constraints on both the layout and components of names and on the structural and signing relationships between names; its statements simply state facts that can be given in any order. The language can only produce entities (Publications, certificates, signing chains, etc.) with a fixed layout so the run-time validatation is loop- and recursion-free. All alternatives (e.g., different signing chains for the same Publication) must be explicitly enumerated so validation and construction are constant time and space. The language is intended for trust domains with a common, local trust anchor. This means the set of signing chains for any schema must be a directed acyclic grap (DAG). This property is verified at compile time and exploited to make the runtime testing for signing-loops stateless and O(1).¶

#mypub: /_domain/_topic/param
req: #mypub & {_topic: "request"}

#mypub & {_topic: "req", param: "status"} | {_topic: "cmd", param: "start"}

#mypub & {_topic: "req"|"cmd", param: "status"|"start"}

cmd: #mypub & {_topic: "cmd"} <= opCert
req: #mypub & {_topic: "req"} <= opCert | userCert

roleCert: /_domain/_role/_roleID
opCert:   roleCert & {_role: "operator"}
userCert: roleCert & {_role: "user"}

#mypub: /_domain/_topic/_roleID/param

4.3. Schema examples

(Figure 15) shows a simple communication schema that defines the format of all Publications in this domain's msgs collection to be an "#mpub", a six component name. The strings between the slashes define the tags used to reference components in the VerSec rules and in the run-time schema library. For example, the tag of the final component is _ts and the & {_ts: timestamp()} constraint on the name makes the component's TLV type a Timestamp (Section 3.3.2) which will be set to the current time when an #mpub is built and be validated as a timepoint when received.¶

The first component gets its value from the variable "_domain" and #pubPrefix is set as this value so that the schema contains information on what part of the name is considered common prefix. (This is a common prefix for all Publications in this domain, not just those in msgs.) The schema puts no constraints on other name components of an #mpub but requires that #mpub Publications be signed by (<=) a mbrCert. Thus the schema enforces that only enrolled domain members can publish.¶

The Validator lines specify cryptographic signing/validation methods for an #mpub, the domain certs, and the PDU that carries Publications. Here all use EdDSA signing.¶

In operation, this schema puts no constraints on #mpubs inner four name components (additional constraints could be imposed by the application but they won't be enforced by DeftT). When it starts, a DeftT adds its signing certificate chain to the domain certificate collection (see Section 5.2), thus announcing its identity to all other members. Using the preconfigured trust anchor and schema, any member can verify the identity of any other member. This approach means members are not preconfigured with identities of other members and new entities can join at any time. Member signing certificates must adhere to the schema; Publications or PDUs with unknown signers are discarded. The timestamp component is used to prevent replay attacks.¶

 #mpub: /_domain/trgt/topic/loc/arg/_ts & { _ts: timestamp() } <= mbrCert

 #pubPrefix:     _domain

 mbrCert:        _domain/_mbrType/_mbrId/_certinfo <= netCert
 netCert:        _domain/_certinfo

 #msgsValidator: "EdDSA"
 #certValidator: "EdDSA"
 #pduValidator:  "EdDSA"

 _domain:        "example"
 _certinfo:      "KEY"/_/"dct"/_

This schema cryptographically identifies the source of every publication and ensures that only enrolled members can publish but offers no privacy beyond that supplied by the lower level communications media. DeftT includes an AEGIS AEAD validator subsystem that leverages each subnet's certificate collection to securely distribute symmetric-encryption nonce cover keys to all members on the subnet. This results in PDU privacy and integrity equivalent to TLS-1.3 at zero-cost (Table 4 measurements show that AEGIS encrypts/decrypts at memory speed).¶

Two changes are required to the above schema to enable this. First, #pduValidator should be changed to "AEAD" instead of "EdDSA". Second, the definition for a keymaker capability cert should be added and made an alternative signer of mbrCerts:¶

kmCap:    _domain/"CAP"/"KM"/_capArg/_certinfo <= netCert
mbrCert:  _domain/_mbrType/_mbrId/_certinfo <= kmCap | netCert

If a schema requires that msgs PDUs be encrypted, all subnet members need to use the same randomly-generated encrypt/decrypt cover key to communicate so DeftT automatically creates a local group key distributor (Section 3.6.2) to obtain it. When each member starts, this distributor checks its group-key collection to see if a keymaker has been elected. If not and if the member's signing chain contains a KM capability, it will start/join a Paxos-like election by publishing a candidate message to the group-key collection. When a keymaker has been elected, members publish requests for the current cover key and the keymaker publishes replies that contain the group key encrypted using the requesting member's public signing key. The keymaker retrieves the member's public signing key from its local store of validated certificates by using the key locator of that member's signed request.¶

In addition to certificate capabilities like KM, VerSec and DeftT allow for Attribute-Based Access Control (ABAC) [NIST] using attributes of the subject, object, action and environment to determine authorization. ABAC takes advantage of the system's context to offer the fine-grained control of capability enumeration but with much less configuration and DeftT uses the schema to manage the ABAC locally at each member. For example, returning to the Figure 15 schema, define the members as either sensors (whose job is to report measurements), loggers (whose job is to solicit then archive sensor reports) or admins (whose job is to configure sensors and loggers). This has the communication pattern that loggers publish request messages to solicit sensor status messages and admins publish config messages to sensors and loggers who reply with state messages. Six rules have to be added to the example to express and enforce this communication pattern: three to distinguish the three member types and three to describe what they can say.¶

sensorCert: mbrCert & { _mbrType: "sensor" } <= netCert
loggerCert: mbrCert & { _mbrType: "logger" } <= kmCap | netCert
adminCert:  mbrCert & { _mbrType: "admin" } <= netCert

sensorMsg: #mpub & {topic:"status"|"state"} <= sensorCert
loggerMsg: #mpub & {topic:"request"|"state"} <= loggerCert
adminMsg:  #mpub & {topic:"config"} <= adminCert

An AEAD kmCap was included in the above to show that ABAC and capabilities are not antagonists. In this case, sensors have limited processors and are often battery powered and admins are people who come and go. Since keymaker is a critical, always on infrastructure service, loggers are the only members qualified to be keymakers but whether any particular logger gets a kmCap is the system administrator's decision which is what the <=kmCap|netCert signing constraint says.¶

Communication schema changes are the only changes required in order to increase security and add privacy: neither code nor binary needs to change and DeftT handles all aspects of validators. This unique approach of integrating identity and communication rules into the transport makes it relatively to produce and evolve secure systems.¶

Converting desired behavioral structure into a schema is the major task of applying Defined-trust Communications to an application domain. Once completed, all the deployment information is contained in the schema. Although a particular schema cert defines a particular trust domain, the text version of a schema can be re-used for related applications. For example, a home IoT schema could be edited to be specific to a particular home network or a solar rooftop neighborhood and then signed with a chosen trust anchor.¶

5.1. Obviate CA usage

Use of third party certificate authorities (CAs) is often antithetical to OT security needs. Any use of a CA (remote or local) results in a single point of failure that greatly reduces system reliability. An architecture with a single, local, trust root cert (trust anchor) and no CAs simplifies trust management and avoids the well-known CA federation and delegation issues and other weaknesses of the X.509 architecture (summarized at [W509], original references include [RSK][NVR]). DCT certificates (see Section 3.3.1.4) can be generated and signed locally (using supplied utilities) so there is no reason to aggregate a plethora of unrelated claims into one cert (avoiding the Aggregation problem [W509]).¶

A DCT cert's one and only Subject Name is the Name of the Publication that contains the public key as its content and neither name nor content are allowed to contain any optional information or extensions. Certificates are created with a lifetime; local production means cert lifetimes can be just as long as necessary (as recommended in [RFC2693]) so there's no need for the code burden and increased attack surface associated with certificate revocation lists (CRLs) or use of on-line certificate status protocol (OSCP). Keys that require longer lifetimes, like device keys, get new certs before the current ones expire and may be distributed through DeftT (e.g., using a variant of the group key distributors in DCT). If there is a need to exclude previously authorized identities from a domain, there are a variety of options. The most expedient is via use of an AEAD cAdd or Publication validator by ensuring that the group key maker(s) of a domain exclude that entity from subsequent symmetric key distributions until its identity cert expires (and it is not issued an update). Another option is to publish an identity that supplants that of the excluded member. A new schema can be distributed at any time but the receiving application will need to start a new DeftT using that schema though it can run in parallel with the old DeftT for some period of time. Though more complex, it is also possible to distribute a new schema and identities (without changing the trust anchor), e.g., using remote attestation via the TPM.¶

From Section 4, a member is granted attributes in the schema via the certs that appear in its identity chain. Identity certs are always accompanied by their full chain-of-trust, both when installed and when the member publishes its identity to the cert collection. Every signing chain in the domain has the same trust anchor at its root and its compliant format specified in the schema. Without the entire chain, a signer's right to issue Publications cannot be validated. Cert validation is according to the schema which may specify attributes and capabilities for Publication signing from any certificate in the chain. For this model to be well founded, each cert's key locator must uniquely identify the cert that actually signed it. This property ensures that each locator resolves to one and only one signing chain. A cert's key locator is a thumbprint, a SHA256 hash of the entire signer's Publication (name, content, key locator, and signature), ensuring that each locator resolves to one and only one cert and signing chain. Use of the thumbprint locator ensures that certs are not open to the substitution attacks of name-based locators like X.509's "Authority Key Identifier" and "Issuer" [ConfusedDep][CAvuln][TLSvuln].¶

5.2. Identity bundles

Identity bundles comprise the certificates needed to participate in a trust domain: trust anchor, schema, and an identity chain. The private key corresponding to the leaf certificate of the identity chain should be installed securely when a device is first commissioned (e.g., out-of-band) for a network while the public certs of the bundle may be placed in a file in a well-known location or may, in addition, have their integrity attested or even be encrypted.¶

Figure 16 shows the steps involved in creating identity bundles and their correspondence of the steps to the "building plans" model. (The corresponding tools available in DCT are shown across the bottom and the relationship to the "building plans" model is shown across the top.)¶

The process in Figure 16 can be used to construct an identity bundle using the example schema from Section 4.2 for a sensor with mbrId = 1. Then the bundle can be inspected using the ls_bundle utility at [DCT] showing the certs it contains. Left hand notations designate the certs by their position in the bundle and show which cert precedes it ("<=") in the chain. The private key for cert 3 is also in this bundle, indicated by "key" at the first bytes of the key value.¶

0 root: /example/KEY/^4c29d555/dct/23-09-20@16:36
1 <= 0: /example/schema/#mpub/KEY/^545b55ff/dct/23-09-20@16:36
2 <= 0: /example/CAP/KM/1/KEY/^77fe722e/dct/23-09-20@16:36
3 <= 2: /example/sensor/1/KEY/^d0660cc1/dct/23-09-20@16:36 key '6ca45831...

In the examples at [DCT], an identity bundle (with private key) is given directly to an application via the command line, useful for development, and the application passes callbacks to utility functions that supply the certs and a signing pair separately. For deployment, good key hygiene using best current practices must be followed e.g., [COMIS]. [DCT] examples use configured member identities to sign locally created signing certs (with associated private keys) and the example schemas give the format for these signing cert names. A DeftT will request a new signing cert shortly before expiration of the one in use. Upon each signing cert update, only this new cert needs to be published via DeftT's cert distributor.¶

In deployment, a small application manager may be programmed for two specific purposes. First, it is registered with a supervisor [SPRV] (or similar process control) for its own (re)start to serve as a bootstrap for the application. Second, it can have access to the TPM functions and the ability to create "short-lived" (hours to several days) public/private key pair(s) that are signed by the installed (commissioned) private identity key using the TPM. This signing key pair is created at (re)start and recreated at the periodicity of the signing cert lifetime. Figure 17 outlines a representative procedure. Since the signing of the public cert happens via requests to the TPM, the identity key (used to sign the cert) cannot be exfiltrated. The locally created signing key is used in the communications path where TPM signing overhead is prohibitive.¶

All DCT certs have a validity period. Publication signing key pairs (with public signing certs) are generated locally so they can easily be refreshed as needed. Trust anchors, schemas, and the member identity chain are higher value and often require generation under hermetic conditions by some authority central to the organization. Their lifetime should be application- and deployment-specific, but the higher difficulty of cert production and distribution often necessitates liftetimes of weeks to years.¶

Secure device configuration and on-boarding should be carried out using the best practices most applicable to a particular deployment. The process of enrolling a device by provisioning an initial secret and identity in the form of public-private key pair and using this information to securely onboard a device to a network has a long history. Current and emergent industry best practices provide a range of approaches for both secure installation and update of private keys. For example, the private key of the bundle can be secured using the Trusted Platform Module, the best current practice in IoT [TATT][DMR][IAWS][TPM][OTPM][SIOT][QTPM][SKH][RFC8995], or secure enclave or trusted execution environment (TEE) [ATZ] where available. In that case, an authorized configurer adding a new device can use TPM tools to secure the private signing key and install the rest of the bundle file in a known location before deploying the device in the network. Where entities have public-private key pair identities of any (e.g., non-DCT) type, these can be leveraged for DeftT identity installation.¶

Updating schemas and other certificates over the deployed network (OTA) is application-domain specific and can either make use of domain best practices or develop custom DeftT-based distribution. Changing the trust anchor is considered a re-commissioning. The example here is merely illustrative; with pre-established secure identities and well-founded approaches to secure on-line communications, a trust domain could be created OTA using secure identities established through some other system of identity.¶

6.1. Secure Industrial IoT

IIoT sensors offer significant advantages in industrial process control including improved accuracy, process optimization, predictive maintenance and analysis, higher efficiency, low-cost remote accessibility and monitoring, reduced downtime, power savings, and reduced costs [IIOT]. Critical Digital Assets (CDA) are a class of industrial assets such as power plants or chemical factories which must be carefully controlled to avoid loss-of-life accidents and where IIoT sensors require tight security. Even when IIoT sensors are not used for direct control of CDA, spoofed sensor readings can lead to destructive behavior. There are real-life examples (such as uranium centrifuges) of nation-state actors changing sensor readings through cyberattacks leading to equipment damage. These risks result in a requirement for stringent security reviews and regulation of CDA sensor networks. Despite the advantages of deploying CDA sensors, adequate security is prerequisite to deploying the CDA sensors. Information conveyed via DeftT has an ensured provenance and may be efficiently encrypted making it ideal for this use.¶

6.2. Secure access to Distributed Energy Resources (DER)

The electrical power grid is evolving to encompass many smaller generators with complex interconnections. Renewable energy systems such as smaller-scale wind and solar generator sites must be economically accessed by multiple users such as building owners, renewable asset aggregators, utilities, and maintenance personnel with varying levels of access rights. North American Electric Reliability Corporation Critical Infrastructure Protection (NERC CIP) regulations specify requirements for communications security and reliability to guard against grid outages [DER]. Legacy NERC CIP compliant utility communications approaches, using dedicated physically secured links to a few large generators, are no longer practical. DeftT offers multiple advantages over bilateral TLS sessions for this use case:¶

• Security. Encryption, authentication, and authorization of all information objects. Secure brokerless pub/sub avoids single-point broker vulnerabilities. Large generation assets of hundreds of megawatts to more than 1 gigawatt, particularly nuclear power plants must be controlled securely or risk large-scale loss of life accidents. Hence, they are attractive targets for sophisticated nation-state cyber attackers seeking damage with national security implications. Even small-scale DER generators are susceptible to a coordinated attack which could still bring down the electric grid.¶
• Scalability. Provisioning, maintaining, and distributing multiple keys with descriptive, institutionalized, hierarchical names. DeftT allows keys to be published and securely updated on-line. Where historically a few hundred large-scale generators could supply all of the energy needs for a wide geographic area, now small-scale DER such as residential solar photovoltaic (PV) systems are located at hundreds of thousands of geographically dispersed sites. Many new systems are added daily and must be accommodated economically to spur wider adoption.¶
• Resiliency. A mesh network of multiple client users, redundant servers, and end devices adds reliability without sacrificing security. Generation assets must be kept on-line continuously or failures risk causing a grid-wide blackout. Climate change is driving frequent natural disasters including wildfires, hurricanes, and temperature extremes which can impact the communications infrastructure. If the network is not resilient communications breakdowns can disable generators on the grid leading to blackouts.¶
• Efficiency. Data can be published once from edge gateways over expensive cellular links and be accessed through servers by multiple authorized users, without sacrificing security. For small residential DER systems, economical but reliable connectivity is required to spur adoption of PV compared to purchasing from the grid. However, for analytics, maintenance and grid control purposes, regular updates from the site by multiple users are required. Pub/sub via DeftT allows both goals to be met efficiently.¶
• Flexible Trust rules: Varying levels of permissions are possible on a user-by-user and site-by-site basis to tightly control user security and privacy at the information object level. In an energy ecosystem with many DER, access requirements are quite complex. For example, a PV and battery storage system can be monitored on a regular basis by a homeowner. Separate equipment vendors for batteries and solar generation assets, including inverters, need to perform firmware updates or to monitor that the equipment is operating correctly for maintenance and warranty purposes. DER aggregators may contract with a utility to supply and control multiple DER systems, while the utility may want to access production data and perform some controls themselves such as during a fire event where the system must be shut down. Different permissions are required for each user. For example, hourly usage data which gives detailed insight into customer behaviors can be seen by the homeowner, but for privacy reasons might only be shared with the aggregator if permission is given. These roles and permissions can be expressed in the communication rules and then secured by DeftT's use of compiled schemas.¶

The specificity of the requirements of NERC CIP can be used to create communication schemas that contain site-specifics, allowing applications to be streamlined and generic for their functionality, rather than containing security and site-specifics.¶