Breaking Deniability in Signal with SGX [pdf]
aaltodoc.aalto.fi
aaltodoc.aalto.fi
I'm struggling to think of a single case, ever, where someone embroiled in a public dispute denied the contents of conversation logs, even in cases where it'd be trivial to do so because the logs were just screenshots of WhatsApp or copy/pasted emails. People have deniable communications already in so many cases and yet nobody cares, nobody actually ever claims "I didn't send that".
To take the Bezos case, he could have just let them leak some dick pics and then said, "not my cock, not my problem". How would anyone have known otherwise?
The only case I can think of where cryptographic proof was used to prevent deniability were some Wikileaks email dumps where there was some very small amount of skepticism that all the emails were real, that maybe some forged emails had been put amongst the real set. The messages were all DKIM signed though, which proved it wasn't true. Which in this case would appear to have been beneficial - the leaks revealed various political shenanigans that were of a public interest nature. But I don't think DKIM ever made it to public consciousness: journalists don't know about or don't care about cryptographic proof.
Now maybe none of that matters, maybe deniability is near-free to implement so why not? But this paper suggests it's more complex than it looks. Any auditable generic computing infrastructure can be used to break deniability. In the unlikely event that disclaiming screenshot-strength evidence becomes a widespread tactic, anyone who wants to defeat it can just use a TPM and trusted boot. Getting deniability back may either seriously complicate already complex implementations (which is the enemy of correctness), or just fail entirely.
Moreover it seems to me that most people, given the chance, would rather than undeniability, as that would be useful in all kinds of regular situations like contract disputes. It'd make casual verbal-but-over-internet-chat type contracts meaningfully useful, along with other interesting capabilities.
Perhaps in the end this protocol feature is a bad tradeoff.
From TFA:
"We then present an practical attack that can be used to compromise deniability in the Signal messaging protocol, and introduce some possible countermeasures that restore deniability
[...]
Deploying the attack presented in this thesis requires not only having access to a TEE but also compromising one of the parties to a communication."
Deniability is a powerful tool which one would rather have and not need than need and not have. Let research run its course for the betterment of protections available to us all.
Essentially everyone with a PC has access to TEEs: beyond Intel SGX there's also the traditional static root of trust mechanism, which suffices for this, and also voids the paper's proposed solution; indeed the paper recognises that given a certain kind of TEE there appear to be no solutions.
As for 'compromising one the parties', the point of deniability implemented cryptographically is that your counterparty in the conversation could encrypt messages such that you can't tell who encrypted them, i.e. an encrypted chat log could have been forged. So the adversary is the person you're talking to: they're already 'compromised'.
Deniability is “i can log into this device to ‘the same’ account two different ways and one of them exposes the super sensitive documents that must be protected and the other merely exposes some mildly rude documents that convince the attackers that i’m not totally bullshitting them”
it’s not a feature most people need in their lives, thankfully.
Now if only Intel could actually build a secure TPM! :P
Clever!