"Stealing" Production Language Models


Posted on in Presentations

This session will discuss two approaches that adversaries use to "steal" information about proprietary black-box AI systems such as ChatGPT. Will first show how providing access to LM output probabilities via API can allow adversaries to reveal the hidden size of LLMs, and will then discuss how training imitation models via distillation can provide a mechanism for creating copycat models.

Access This and Other RSAC Conference Presentations with Your Free RSAC Membership

Your RSAC Membership also includes AI-powered summaries, mind maps, and slides for Conference presentations, Group Discussions with experts, and more.

Watch Now >>
Participants
Saurabh Shintre

Moderator

CEO, Realm Labs

Eric Wallace

Speaker

PhD Student, UC Berkeley


Share With Your Community