Stealing Part of a Production LLM | API protects LLMs no more
How it is possible to steal part of LLMs protected behind an API? 🥷 We explain both papers that made a breakthrough on this, one from Carlini et al. (Google), and the other one from Finlayson et al. (USC), see references below. SPONSOR: AssemblyAI 👉 https://www.assemblyai.com/research/universal-1/?utm_source=youtube&utm_medium=social&utm_campaign=universal1_letitia AI Coffee Break Merch! 🛍️ https://ai-coffee-break-with-letitia.myspreadshop.de/ 📄 Carlini, Nicholas, Daniel Paleka, Krishnamurthy Dj Dvijotham, Thomas Steinke, Jonathan Hayase, A. Feder Cooper, Katherine Lee et al. "Stealing Part of a Production Language Model." https://arxiv.org/abs/2403.06634 📄 Finlayson, Matthew, Swabha Swayamdipta, and Xiang Ren. "Logits of API-Protected LLMs Leak Proprietary Information." https://arxiv.org/abs/2403.09539 Thanks to our Patrons who support us in Tier 2, 3, 4: 🙏 Dres. Trost GbR, Siltax, Vignesh Valliappan, Michael, Sunny Dhiana, Andy Ma Outline: 00:00 Stealing LLMs from behind API’s!? 01:54 AssemblyAI (Sponsor) 02:59 Two papers, same thing 04:03 Core observation 07:05 Recover Hidden Dimensionality 08:54 gpt-3.5-turbo 10:30 Full Layer Extraction 10:58 Extract all logits 14:35 Defenses 15:40 Cost of attack 16:22 Further impact 17:40 API response stochasticity ▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ 🔥 Optionally, pay us a coffee to help with our Coffee Bean production! ☕ Patreon: https://www.patreon.com/AICoffeeBreak Ko-fi: https://ko-fi.com/aicoffeebreak Join this channel to get access to perks: https://www.youtube.com/channel/UCobqgqE4i5Kf7wrxRxhToQA/join ▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ 🔗 Links: AICoffeeBreakQuiz: https://www.youtube.com/c/AICoffeeBreak/community Twitter: https://twitter.com/AICoffeeBreak Reddit: https://www.reddit.com/r/AICoffeeBreak/ YouTube: https://www.youtube.com/AICoffeeBreak #AICoffeeBreak #MsCoffeeBean #MachineLearning #AI #research Video editing: Nils Trost Scientific advising by Mara Popescu