Discussion with the 2026 USENIX Security Symposium Test of Time
summary
Panel traces the 2016 claim that cloud-hosted models can be stolen via prediction API queries, demonstrating practical model extraction attacks that recover counterparts with high fidelity by issuing systematic queries to Google/AWS-style services. Experiments target real deployed services, reveal many offerings are just logistic regression models behind the marketing, and show the attack scales with query budget—sparking follow-up work, high citation counts, and new defenses and threat models.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 1:41Award-winning paper origins
- 3:05Collaboration origins
- 6:58Research evolution
- 10:48Closing remarks
related talks
28mStealing Machine Learning Models via Prediction APIs
45mDiving into the Tenacious Threats of Cloud Data Lakes
33mUnderstanding the Threats and Attacks on Data Science Applications and Models
52mSleeping With One AI Open: An Introduction to Attacks Against...
46mInsane in the Supply Chain: Threat modeling for...
20mAttacking and Defending Machine Learning Applications of Public Cloud
13mHow Cloud AI Leaks What It Learns
24mDiscussion between Mehdi Medjaoui and Troy Leliard,API Security.