Adversarial Distillation: OpenAI’s Model Theft Defense
OpenAI’s analysis of adversarial distillation examines how attackers coax target behaviors from proprietary models and transfer them elsewhere. It also explains how detection, disruption, and tracing defenses could affect developers using the API.