Seminars and Events
NL Seminar- The Search for Unforgeable Language Model Signatures
Event Details
Speaker: Matt Finlayson, USC
Conference Rm Location: ISI-MDR CRs#1135-1137
Zoom Link: https://usc.zoom.us/j/99470400418?pwd=bRgL0U5ICvscL9IaYcab02sBsyCCRx.1
Meeting ID: 994 7040 0418Passcode: 246810
REMINDER:
Meeting hosts only admit on-line guests that they know to the Zoom meeting. Hence, you’re highly encouraged to use your USC account to sign into Zoom.
If you’re an outside visitor, please inform us at (nlg-seminar-host(at)isi.edu) to make us aware of your attendance so we can admit you. Specify if you will attend remotely or in person at least one business day prior to the event. Provide your: full name, job title and professional affiliation and arrive at least 10 minutes before the seminar begins.
Abstract: For security reasons, language model inference providers generally limit access to their models via restrictive blackbox APIs. These limited APIs protect model secrets, but erode trust and accountability, since clients have no guarantees about which model is actually being served. Our work shows the existence of geometric language model signatures, which leverage the geometry of the model’s final layers to identify model-specific constraints on next-token distributions. Checking whether an output came from a specific model requires only checking that the outputs satisfy the model constraints. We additionally develop the notion of signature forgery, where adversaries attempt to copy the geometric signature from a target model into another by analyzing API outputs from the target model. We prove the existence computationally unforgeable signatures, by showing that the geometric constraints on token rankings are NP-hard to forge. Our results add a new suite of tools for language model forensics, with applications in provider accountability and security.