Post a Comment On: Not Just AstLinux Stuff

This is a post that would be received completely differently a year ago.
We now live in a “post-Snowden” era where even average members of the general public are at least vaguely aware of things like “wiretaps” and “encryption”. With that said this isn’t another post about those revelations.

Over the past year I’ve been steadily working to improve the security of the various real-time communications platforms I’m involved in. Timing on this has been coincidental with the news about the NSA and the various programs operated by them. Of course with these revelations I’ve been paying extra special attention to key sizes, cipher suites, and of course the use of elliptic curve cryptography.

As regular readers of my blog might suspect my primary technical focus is on SIP and RTP.  SIP provides for the use of TLS to secure the signaling channel. While TLS isn’t always TLS (TLS versions, cipher suites, etc) there isn’t anything special about the application of TLS to SIP.

TLS is one of those things we just take for granted these days (kids - get off my lawn)!  However, there are several well known issues with various TLS implementations currently in use. Even though TLS is used to secure banking and highly sensitive web traffic everyday, SSL Labs reports that only 20% of web servers currently support TLS 1.2. Yikes.

Of course TLS isn’t just for web traffic. SIP can be secured with TLS as well. This is a very interesting scenario because as the name implies, SIP is almost exclusively used as a signaling channel to establish another session (which may or may not be secured). This means that securing SIP with TLS is not only important for the integrity of the SIP session, it’s often reused to secure the communications established by that session. This is the case with real time voice and video communications using RTP or more specifically the secure profile of RTP (SRTP).

As always, anything SIP related is complicated and security is no exception. To offer a completely encrypted solution the alphabet soup and resulting complexity adds up pretty quickly.  First, you must establish a secure signaling channel using SIP over TLS. You then use SDES (hopefully) to exchange offers and key information over SDP. Once the SDP offer answer model has done it’s thing you will have negotiated the transport of secure voice and/or video using SRTP.

Originally defined in RFC 3711, SRTP has now been updated with several follow-ups, errata, and extensions. Looking around in the open source community and commercial VoIP space, however, shows that many implementations have not moved very far beyond this original specification for various reasons.

One of these reasons is due to existing libraries implementing SRTP. Just about all of the open source implementations of SRTP that I’m aware of use some version of the original libsrtp written by David McGrew of Cisco Systems over seven years ago. Since then libsrtp has been forked, patched, and otherwise used by pjsip, FreeSWITCH, Asterisk, the Chromium/Chrome web browsers from Google (for WebRTC), and countless other projects.

For some time now I’ve been interested in replacing the built in AES and SHA functions in libsrtp with those provided by OpenSSL. This goal led me to the work of John Foley (also of Cisco Systems) and the official libsrtp Github repository. When I first started looking into this John had (recently) completed work on a branch of libsrtp that accomplishes exactly this.  He also went two steps further: the use of AES-NI and support for AES-GCM mode in SRTP.

What?

AES-NI is the reason I was first interested in using OpenSSL for SRTP. Long story short new(ish) Intel CPUs include a new instruction set that greatly increases the performance of AES (especially with AES-GCM and larger payload sizes). This is a good thing and should (hopefully) lead to more widespread adoption of SRTP.

The use of OpenSSL included another perk: support for AES-GCM. While AES-GCM isn’t necessarily new there hadn’t been support for it with SRTP yet. AES-GCM is the current “gold standard” AES block cipher mode of operation, especially for “streaming” operations like VPNs and other packetized data. In fact, Cisco currently considers AES-GCM a “next generation encryption” solution.  AES-GCM is also preferred by the National Security Agency’s (NSA) Suite B crypto specification (for better or worse in a “post-Snowden” world).

AES-GCM is one of those “standing on the shoulders of giants” moments I really appreciate. To think that I’m writing a blog post in 2014 talking about a technology that uses a field named after an early nineteenth century French mathematician (with a really interesting life) is amazing to me. As Isaac Newton said “If I have seen further it is by standing on the shoulders of giants.” In this case the giant was born over 200 years ago and only lived to the age of 20.

Anyway, John Foley (another giant) was able to provide a fork of pjsip that included the feature-openssl branch of libsrtp and had been patched to support AES-GCM mode with SIP (SDP/SDES) and SRTP. He just didn’t have anything else to test it against.

I wrote patches implementing 128 bit AES-GCM mode using the updated libsrtp branch for both FreeSWITCH and Asterisk. Because of limitations in the original SRTP specification both projects would require fairly significant refactoring to include support for 192 and 256 bit GCM modes, “big AES SRTP”, and the use of GCM mode with DTLS (for WebRTC). In fact, both FreeSWITCH and Asterisk implemented support for SRTP by largely differentiating only on authentication tag length.

In any case we were able to use these patches to test 128 bit AES-GCM mode SRTP interoperability between pjsip, Asterisk, and FreeSWITCH. However, there was still work to be done.

I’m happy to report that as of last week the FreeSWITCH master branch now not only has support for the use of 128, 192 and 256 bit AES-GCM modes it also includes a very flexible means to control the use of various SRTP crypto suites on both incoming and outgoing channels. Unfortunately this hasn’t been implemented for DTLS use with WebRTC yet but I plan to advocate for that shortly. I’m also still working with the Asterisk project to clean up my patch for inclusion there.

I think it’s safe to say that with these changes FreeSWITCH now supports state of the art (some may say “next generation”) cryptography with SRTP. However, if you recall the keys used for SRTP are still exchanged using SIP over TLS, which depending on various factors, could still have numerous issues and be fundamentally insecure. How can we be sure that our SIP signaling channel is properly secured?

FreeSWITCH and Kamailio now both also support the explicit specification of TLS version and cipher suite. When using OpenSSL 1.0.1 (which is required by FreeSWITCH) this allows the user to require the use of TLS 1.2 and only the strongest cipher suites supported by OpenSSL. Of course this is great news for FreeSWITCH and Kamailio, but what about the rest of the SIP+SRTP implementations out there?

Unfortunately support for these standards is bleeding edge at this point and not widely deployed. While any use of SIP over TLS and SRTP is certainly preferred for various reasons (performance and security) I’m pushing for support of TLS 1.2 and AES-GCM everywhere possible.

Hopefully I’ll be able to offer more on that soon but for now enjoy the faster and more secure real time communications made possible using FreeSWITCH and Kamailio!