Signaling Additional SRTP Context information via SDP

1.1. Discussion Venues

This section is to be removed before publishing as an RFC.¶

Source for this draft and an issue tracker can be found at https://github.com/kyzer-davis/srtp-assurance-rfc-draft.¶

1.2. Changelog

This section is to be removed before publishing as an RFC.¶

draft-01¶

• Change contact name from IESG to IETF in IANA Considerations #2¶

• Discuss RFC4568 "Late Joiner" in problem statement: #3¶

• Split Serial forking scenario into its own section #4¶

• Add MIKEY considerations to Protocol Design section #6¶

• Change doc title #7¶

• Add SEQ abbreviation earlier #8¶

• Discuss why this can't be a RTP Header Extension #11¶

• Add Appendix further discussing why SDP Security Session Parameters extension not used #5¶

• Method to Convey Multiple SSRCs for a given stream #1¶

• Discuss why SEQ is signaled in the SDP #9¶

1.3. Problem Statement

While [RFC4568] provides most of the information required to instantiate an SRTP cryptographic context for RTP Packets, the state of a few crucial items in the SRTP cryptographic context are missing. One such item is the Rollover Counter (ROC) defined by Section 3.2.1 [RFC3711] which is not signaled in any packet across the wire and shared between applications.¶

The ROC is one item that is used to create the SRTP Packet Index along with the the [RFC3550] transmitted sequence numbers (SEQ) for a given synchronization sources (SSRC). The Packet index is integral to the encryption, decryption and authentication process of SRTP key streams. Failure to synchronize the value properly at any point in the SRTP media exchange leads to encryption or decryption failures, degraded user experience and at cross-vendor interoperability issues with many hours of engineering time spent debugging a value that is never negotiated on the wire (and oftentimes not even logged in application logs.)¶

The current method of ROC handling is to instantiate a new media stream's cryptographic context at 0 as per Section 3.3.1 of [RFC3711]. Then track the state ROC for a given cryptographic context as the time continues on and the stream progresses.¶

[RFC4568], states 'there is no concept of a "late joiner" in SRTP security descriptions' as the main reason for not conveying the ROC, SSRC, or SEQ via the key management protocol but as one will see below; this argument is not true in practice.¶

When joining ongoing streams, resuming held/transferred streams, or devices without embedded application logic for clustering/high availability where a given cryptographic context is resumed; without any explicit signaling about the ROC state, devices must make an educated guess as defined by Section 3.3.1 of [RFC3711]. The method specially estimates the received ROC by calculating ROC-1, ROC, ROC+1 to see which performs a successful decrypt. While this may work on paper, this process usually only done at the initial instantiation of a cryptographic context rather than at later points later during the session. Instead many applications take the easy route and set the value at 0 as if this is a new stream. While technically true from that receivers perspective, the sender of this stream may be encrypting packets with a ROC greater than 0. Further this does not cover scenarios where the ROC is greater than +1.¶

Where possible the ROC state (and the rest of the cryptographic context) is usually synced across clustered devices or high availability pairs via proprietary methods rather than open standards.¶

These problems detailed technically above lead to a few very common scenarios where the ROC may become out of sync. These are are briefly detailed below with the focus on the ROC Value.¶

Joining an ongoing session:¶

• When a receiver joins an ongoing session, such as a broadcast conference, there is no signaling method which can quickly allow the new participant to know the state of the ROC assuming the state of the stream is shared across all participants.¶

Hold/Resume, Transfer Scenarios:¶

• A session is created between sender A and receiver B. ROC is instantiated at 0 normally and continues as expected.¶

• At some point the receiver is put on hold while the sender is connected to some other location such as music on hold or another party altogether.¶

• At some future point the receiver is reconnected to the sender and the original session is resumed.¶

• The sender may re-assume the original cryptographic context rather rather than create one net new.¶

• Here if the sender starts the stream from the last observed sequence number the receiver observed the ROC will be in sync.¶

• However there are scenarios where the sender may have been transmitting packets on the previous cryptographic context and if a ROC increment occurred; the receiver would never know. This can lead to problems when the streams are reconnected as the ROC is now out of sync between both parties.¶

• Further, a sender may be transferred to some upstream device transparently to them. If the sender does not reset their cryptographic context that new receiver will now be out of sync with possible ROC values.¶

Serial Forking Case:¶

• [RFC4568] itself cites a problematic scenario in their own Appendix A, Scenario B, Problem 3 where a ROC out of sync scenario could occur.¶

• The proposed solution for problem 3 involves a method to convey the ROC however known the problem; the authors still did not include this in the base SDP Security specification.¶

Application Failover (without stateful syncs):¶

• In this scenario a cryptographic context was was created with Device A and B of a high availability pair.¶

• An SRTP stream was created and ROC of 0 was created and media streamed from the source towards Device A.¶

• Time continues and the sequence wraps from 65535 to 0 and the ROC is incremented to 1.¶

• Both the sender and device A are tracking this locally and the encrypt/decrypt process proceeds normally.¶

• Unfortunate network conditions arise and Device B must assume sessions of Device A transparently.¶

• Without any proprietary syncing logic between Device A and B which disclose the state of the ROC, Device B will likely instantiate the ROC at 0.¶

• Alternatively Device B may try to renegotiate the stream over the desired signaling protocol however this does not ensure the remote sender will change their cryptographic context and reset the ROC to 0.¶

• The transparent nature of the upstream failover means the local application will likely proceed using ROC 1 while upstream receiver has no method of knowing ROC 1 is the current value.¶

Secure SIPREC Recording:¶

• If a SIPREC recorder is brought into recording an ongoing session through some form of transfer or on-demand recording solution the ROC may have incremented.¶

• Without an SDP mechanism to share this information the SIPREC will be unaware of the full SRTP context required to ensure proper decrypt of media streams being monitored.¶

Improper SRTP context resets:¶

• As defined by Section 3.3.1 of [RFC3711] an SRTP re-key MUST NOT reset the ROC within SRTP Cryptographic context.¶

• However, some applications may incorrectly use the re-key event as a trigger to reset the ROC leading to out-of-sync encrypt/decrypt operations.¶

Out of Sync Sliding Windows / Sequence Numbers:¶

• There is corner case situation where two devices communicating via a Back to Back User Agent (B2BUA) which is performing RTP-SRTP inter-working.¶

• In this scenario the B2BUA is also a session border controller (SBC) tasked with topology abstraction. That is, the signaling itself is abstracted from both parties.¶

• In this scenario a hold/resume where a sequence rolls can not only cause problems with the ROC; but can also cause sliding window issues.¶

• To be more specific, assume that both parties did have access to the cryptographic context and resumed the old ROC value after the hold thus ROC is not out of sync.¶

• What should the sliding window and sequence be set to in this scenario?¶

• The post-hold call could in theory have a problem where the sequence number of received packets is lower than what was originally observed before the hold.¶

• Thus the sliding window would drop packets until the sequence number gets back to the last known sequence and the sliding window advances.¶

• Advertising the Sequence in some capacity to reinitialize the sliding window (along with advertising the ROC) can ensure a remote application can properly re-instantiate the cryptographic context in this scenario.¶

This is a problem that other SRTP Key Management protocols (MIKEY, DTLS-SRTP, EKT-SRTP) have solved but SDP Security has lagged behind in solution parity. For a quick comparison of all SRTP Key Management negotiations refer to [RFC7201] and [RFC5479].¶

1.4. Previous Solutions

As per RFC3711, "Receivers joining an on-going session MUST be given the current ROC value using out-of-band signaling such as key-management signaling." [RFC4771] aimed to solve the problem however this solution has a few technical shortcomings detailed below.¶

First, this specifies the use of Multimedia Internet KEYing (MIKEY) defined by [RFC3830] as the out-of-band signaling method. A proper MIKEY implementation requires more overhead than is needed to convey and solve this problem. By selecting MIKEY as the out-of-band signaling method the authors may have inadvertently inhibited significant adoption by the industry.¶

Second, [RFC4771] also transforms the SRTP Packet to include the four byte value after the encrypted payload and before an optional authentication tag. This data about the SRTP context is unencrypted on the wire and not covered by newer SRTP encryption protocols such as [RFC6904] and [RFC9335]. Furthermore this makes the approach incompatible with AEAD SRTP Cipher Suites which state that trimming/truncating the authentication tag weakens the security of the protocol in Section 13.2 of [RFC7714].¶

Third, this is not in line with the standard method of RTP Packet modifications. The proposal would have benefited greatly from being an RTP Header Extension rather than a value appended after payload. But even an RTP header extension proves problematic in where modern SRTP encryption such as Cryptex defined by [RFC9335] are applied. That is, the ROC is a required input to decrypt the RTP packet contents. It does not make sense to convey this data as an RTP Header Extension obfuscated by the very encryption it is required to decrypt.¶

Lastly, there is no defined method for applications defined for applications to advertise the usage of this protocol via any signaling methods.¶

[RFC5159] also defined some SDP attributes namely the "a=SRTPROCTxRate" attribute however this does not cover other important values in the SRTP Cryptographic context and has not seen widespread implementation.¶

[RFC8870] solves the problem for DTLS-SRTP [RFC5763]/[RFC5764] implementations.¶

3.1. SDP Considerations

This specification introduces a new SRTP Context attribute defined as "a=srtpctx".¶

The presence of the "a=srtpctx" attribute in the SDP (in either an offer or an answer) indicates that the endpoint is signaling explicit cryptographic context information and this data SHOULD be used in place of derived values such as those obtained from late binding or some other mechanism.¶

The SRTP Context value syntax utilizes standard attribute field=value pairs separated by semi-colons as seen in Figure 1. The implementation's goal is extendable allowing for additional vendor specific field=value pairs alongside the ones defined in this document or room for future specifications to add additional field=value pairs.¶

a=srtpctx:<a-crypto-tag> \
  <att_field_1>=<value_1>;<att_field_1>=<att_value_2>

This specification specifically defines SRTP Context Attribute Fields of SSRC, ROC, and SEQ shown in Figure 2.¶

a=srtpctx:<a-crypto-tag> \
  ssrc=<ssrc_value_hex>;roc=<roc_value_hex>;seq=<last_known_tx_seq_hex>

Note that long lines in this document have been broken into multiple lines using the "The Single Backslash Strategy ('')" defined by [RFC8792].¶

The formal definition of the SRTP Context Attribute, including custom extension field=value pairs is provided by the following ABNF [RFC5234]:¶

srtp-context   = srtp-attr
                 srtp-tag
                 [srtp-ssrc";"]
                 [srtp-roc";"]
                 [srtp-seq";"]
                 [srtp-ext";"]
srtp-attr      = "a=srtpctx:"
srtp-tag       = 1*9DIGIT 1WSP
srtp-ssrc      = "ssrc=" ("0x"1*8HEXDIG / "unknown")
srtp-roc       = "roc=" ("0x"1*4HEXDIG / "unknown")
srtp-seq       = "seq=" ("0x"1*4HEXDIG / "unknown")
srtp-ext       = 1*VCHAR "=" (1*VCHAR / "unknown")
ALPHA          = %x41-5A / %x61-7A   ; A-Z / a-z
DIGIT          = %x30-39
HEXDIG         = DIGIT / "A" / "B" / "C" / "D" / "E" / "F"
VCHAR          = %x21-7E

Leading 0s may be omitted and the alphanumeric hex may be upper or lowercase but at least one 0 must be present. Additionally the "0x" provided additional context that these values are hex and not integers. Thus as per Figure 3 these two lines are functionally identical:¶

a=srtpctx:1 ssrc=0x00845FED;roc=0x00000000;seq=0x005D
a=srtpctx:1 ssrc=0x845fed;roc=0x0;seq=0x05d

When SSRC, ROC, or Sequence information needs to be conveyed about a given stream, the a=srtpctx attribute is coupled with the relevant a=crypto attribute in the SDP.¶

In Figure 4 the sender has shares the usual cryptographic information as per a=crypto but has included other information such as the 32 bit SSRC, 32 bit ROC, and 16 bit Last Known Sequence number as Hex values within the a=srtpctx attribute. Together these two attributes provide better insights as to the state of the SRTP cryptographic context from the senders perspective.¶

a=crypto:1 AEAD_AES_256_GCM \
  inline:3/sxOxrbg3CVDrxeaNs91Vle+wW1RvT/zJWTCUNP1i6L45S9qcstjBv+eo0=\
  |2^20|1:32
a=srtpctx:1 ssrc=0x00845FED;roc=0x0000;seq=0x0150

The value of "unknown" MAY be used in place of any of the fields to indicate default behavior SHOULD be utilized by the receiving application (usually falling back to late binding or locally derived/stored cryptographic contact information for the packet index.) The example shown in Figure 5 indicates that only the SSRC of the stream is unknown to the sender at the time of the SDP exchange but values for ROC and Last Known Sequence are present. Alternatively, the attribute key and value MAY be omitted entirely.¶

This MAY be updated via signaling at any later time but applications SHOULD ensure any offer/answer has the appropriate SRTP Context attribute.¶

Applications SHOULD NOT include SRTP Context attribute if all three values are unknown or would be omitted. For example, starting a new sending session instantiation or for advertising potential cryptographic attributes that are part of a new offer.¶

Figure 5 shows that tag 1 does not have any SRTP Context parameters rather than rather an SRTP Context attribute with all three values set to "unknown". This same example shows an unknown value carried with tag 2 and seq has been committed leaving only the ROC as a value shared with the second a=crypto tag.¶

a=crypto:1 AES_CM_128_HMAC_SHA1_32 \
  inline:k4x3YXkTD1TWlNL3BZpESzOFuxkBZmTo0vGa1omW
a=crypto:2 AES_CM_128_HMAC_SHA1_80 \
  inline:PS1uQCVeeCFCanVmcjkpPywjNWhcYD0mXXtxaVBR
a=srtpctx:2 ssrc=unknown;roc=0x0001

The tag for an SRTP Context attribute MUST follow the peer SDP Security a=crypto tag for a given media stream (m=). The example in shown in Figure 6 the sender is advertising an explicit packet index mapping for a=crypto tag 2 for the audio stream and tag 1 for the video media stream. Note that some SDP values have been truncated for the sake of simplicity.¶

c=IN IP4 192.0.0.1
m=audio 49170 RTP/SAVP 0
a=crypto:1 AES_CM_128_HMAC_SHA1_80 \
  inline:d0RmdmcmVCspeEc3QGZiNWpVLFJhQX1cfHAwJSoj|2^20|1:32
a=crypto:2 AEAD_AES_256_GCM \
  inline:HGAPy4Cedy/qumbZvpuCZSVT7rNDk8vG4TdUXp5hkyWqJCqiLRGab0KJy1g=
a=srtpctx:2 ssrc=0xBFBDD;roc=0x0001;seq=0x3039
m=video 49172 RTP/SAVP 126
a=crypto:1 AEAD_AES_128_GCM \
  inline:bQJXGzEPXJPClrd78xwALdaZDs/dLttBLfLE5Q==
a=srtpctx:1 ssrc=0xDD147C14;roc=0x0001;seq=0x3039

It is unlikely a sender will send SRTP Context attributes for every crypto attribute since many will be fully unknown (such as the start of a session.) However it is theoretically possible for every a=crypto tag to have a similar a=srtpctx attribute for additional details.¶

For scenarios where RTP Multiplexing are concerned, EKT-SRTP ([RFC8870]) SHOULD be used in lieu of SDP Security as per [RFC8872] Section 4.3.2. If SRTP Context attributes are to be used, multiple SSRC/ROC/SEQ values can be "bundled" in a list using parenthesis as a delimter. This can be observed in Figure 7 where three SSRC and the respective ROC/SEQ are provided as a list within the a=srtpctx attribute:¶

a=crypto:1 AES_CM_128_HMAC_SHA1_80 \
inline:d0RmdmcmVCspeEc3QGZiNWpVLFJhQX1cfHAwJSo
a=srtpctx:1 (ssrc=0x01;roc=0x0;seq=0x1234), \
(ssrc=0x02;roc=0x1;seq=0xABCD), \
(ssrc=0x845fed;roc=0x0000;seq=unknown)

For scenarios where SDP Bundling are concerned, SRTP Context attributes follow the same bundling guidelines defined by [RFC8859], section 5.7 for SDP Securities a=crypto attribute.¶

3.2. Sender Behavior

Senders utilizing SDP Security via "a=crypto" MUST make an attempt to signal any known packet index values to the peer receiver. The exception being when all values are unknown, such as at the very start of medias stream negotiation.¶

For best results all sending parties of a given session stream SHOULD advertise known packet index values for all media streams. This should continue throughout the life of the session to ensure any errors or out of sync errors can be quickly corrected via new signaling methods. See Section 3.4 for update frequency recommendations.¶

3.3. Receiver Behavior

Receivers SHOULD utilize the signaled information in application logic to instantiate the SRTP cryptographic context. In the even there is no SRTP Context attributes present in SDP receivers MUST fallback to [RFC3711] for guesting the ROC and [RFC4568] logic for late binding to gleam the SSRC and sequence numbers (SEQ).¶

3.4. Update Frequency

Senders SHOULD provide SRTP Context SDP when SDP Crypto attributes are negotiated. There is no explicit time or total number of packets in which a new update is required from sender to receiver. This specification will not cause overcrowding on the session establishment protocol's signaling channel if natural session updates, session changes, and session liveliness checks are followed.¶

3.5. Extendability

As stated in Section 3.1, the SRTP Context SDP implementation's goal is extendability allowing for additional vendor specific field=value pairs alongside the ones defined in this document. This ensures that a=crypto SDP security may remain compatible with future algorithms that need to signal cryptographic context information outside of what is currently specified in [RFC4568].¶

To illustrate, imagine a new example SRTP algorithm and crypto suite is created named "FOO_CHACHA20_POLY1305_SHA256" and the application needs to signal "Foo, "Bar", and "Nonce" values to properly instantiate the SRTP context. Rather than modify a=crypto SDP security or create a new unique SDP attribute, one can simply utilize SRTP Context SDP's key=value pairs to convey the information. Implementations MUST define how to handle default scenarios where the value is not present or set to "unknown".¶

a=crypto:1 FOO_CHACHA20_POLY1305_SHA256 \
 inline:1ef9a49f1f68f75f95feca6898921db8c73bfa53e71e33726c4c983069dd7d44
a=srtpctx:1 foo=1;bar=abc123;nonce=8675309

With this extendable method, all that is now required in the fictional RFC defining "FOO_CHACHA20_POLY1305_SHA256" is to include an "SDP parameters" section which details the expected "a=srtpctx" values and their usages. This approach is similar to how Media Format Parameter Capability ("a=fmtp") is utilized in modern SDP. An example is [RFC6184], Section 8.2.1 for H.264 video Media Format Parameters.¶

A.1. Why not an RTP Header Extension?

In order to be compatible with "a=cryptex", a protocol which extends the SRTP encryption over the RTP Extension Headers, the designed specification must ensure that information about the SRTP context is not within these RTP extension headers. Thus one has to provide this information in an out of band mechanism.¶

A.2. Why not an SDP Security Session Parameter?

While analyzing SDP Security's Session Parameter feature number of interesting details were found. That is sections 6.3.7, 7.1.1, 9.2, and 10.3.2.2 of [RFC4568] specifically.¶

A few illustrative examples below detail what this could look like are provided below, though these MUST NOT be used.¶

a=crypto:1 [..omitted..] SSRC=0x00845FED ROC=0x00000000 SEQ=0x005D

a=crypto:1 ..omitted.. -SSRC=0x00845FED -ROC=0x00000000 -SEQ=0x005D

a=crypto:1 AEAD_AES_256_GCM \
 inline:3/sxOxrbg3CVDrxeaNs91Vle+wW1RvT/zJWTCUNP1i6L45S9qcstjBv+eo0=\
 |2^20|1:32 SSRC=0x00845FED ROC=0x0000 SEQ=0x0150

a=crypto:1 AES_CM_128_HMAC_SHA1_80 \
  inline:QUJjZGVmMTIzNDU2Nzg5QUJDREUwMTIzNDU2Nzg5|2:18\
  ;inline:QUJjZGVmMTIzNDU2Nzg5QUJDREUwMTIzNDU2Nzg5|21|3:4 \
  KDR=23 FEC_ORDER=SRTP_FEC UNENCRYPTED_SRTP \
  SSRC=0xDD148F16 ROC=0x0 SEQ=0x5A53
a=crypto:2 AES_CM_128_HMAC_SHA1_32 \
  inline:QUJjZGVmMTIzNDU2Nzg5QUJDREUwMTIzNDU2Nzg5|2^20 \
  FEC_KEY=inline:QUJjZGVmMTIzNDU2Nzg5QUJDREUwMTIzNDU2Nzg5|2^20|2:4 \
  WSH=60 SSRC=0xD903 ROC=0x0002 SEQ=0xB043
a=crypto:3 AEAD_AES_256_GCM \
  inline:HGAPy4Cedy/qumbZvpuCZSVT7rNDk8vG4TdUXp5hkyWqJCqiLRGab0KJy1g= \
  UNAUTHENTICATED_SRTP SSRC=0x05 ROC=0x02 SEQ=unknown
a=crypto:4 AEAD_AES_128_GCM \
  inline:bQJXGzEPXJPClrd78xwALdaZDs/dLttBLfLE5Q== \
  UNENCRYPTED_SRTCP SSRC=0x6500

To analyze the faults of this method: First, a unknown and/or unsupported SDP Security Session Parameter is destructive. If one side where to advertise the ROC value as an SDP Security Session Parameter and the remote party does not understand that specific SDP Security Session Parameter, that entire crypto line is to be considered invalid. If this is the only a=crypto entry then the entire session may fail. The solution in this document allows for a graceful fallback to known methods to determine these value. Implementations could get around this by duplicating the a=crypto SDP attribute into two values: one with the postfix and one without to create to potential offers; but at this point we have a second SDP attribute. Instead this specification decided to cut to the chase and format the second attribute in a standardized way and avoid endless duplication (and potentially other harmful issues, see the final item in this document.)¶

Second, there is a method to advertise "optional" SDP Security Session Parameters. However, upon further scrutiny, the document contradicts itself in many sections. To be specific, Section 6.3.7 states that an SDP Security Session Parameter prefixed with a dash character "-" MAY be ignored. Subsequent sections (9.2 and 10.3.2.2) state that a dash character is illegal and MUST NOT be used. It is not very well defined as such pursuit of this method has been dropped.¶

Further, we know how applications will handle unknown SDP attributes; we do not know how applications will handle new mandatory (or optional) SDP Security Session Parameter values as none have ever been created. See IANA registry which only details those from the original RFC. (https://www.iana.org/assignments/sdp-security-descriptions/sdp-security-descriptions.xhtml#sdp-security-descriptions-4) Including these could cause larger application issues and are the reason modern protocols use logic like Generate Random Extensions And Sustain Extensibility (GREASE) to catch bad implementation behavior and correct it before it leads to problems like those described in this section.¶

In closing, this method has too many challenges but a lot has been learned. These items have influenced the protocol design and sections like Section 3.5 which aim to avoid making the same mistakes.¶