Mitigating deep fakes and phishing in meetings
Here comes a quick project idea. It’s meant to provide deep-fake + phishing resistance while at the same time not introduce “big bang changes” to how you operate. This allows you to retrofit a system like this into an existing legacy system.
The setting is this:
- you run security sensitive meetings with a bunch of participants. say you run those using a laptop over (say) google meet / zoom / whatever.
- you are remote/distributed
- you are concerned about a participant getting pwned and getting replaced by a deep fake
- you can’t change 100% how you operate. you are stuck with google meet / zoom / whatever
The problem:
- you are concerned some participant is getting MITM / phished and replaced by a deep fake
Solution in a nutshell:
- each participant has an iPhone, separate from your corporate laptop
- iPhones are live transcribing what you say + what others say thru the google meet into text
- they encrypt + authenticate this text and send it to a “group chat”
- at the same time, the iphone is monitoring this group chat and detecting whether other iphones are saying in the group chat actually matches its own transcription.
- if they are consistent, the screen of the iPhone is green. If they are not, a deep fake is detected and it should be red and complain loudly
Product notes:
- This is invisible in case everything goes right
- This is loud in case security issues are detected
Implementation details:
- preferably, this is a personal iPhone separate from corporate infrastructure
- all transcription + diarization can happen locally, say with openAI’s whisper models
- iPhones can use secure enclaves for storing keys
- iPhones can communicate out of band
- the transcription match won’t be perfect, we need some sort of “approximate match”
- of course this assumes remote-only attacker that can’t get a hold of these iPhones