Generative AI Model Extraction & Data Leakage
Interactive 3D simulator of two real attacks on generative AI APIs: model extraction (cloning a model via input-output query pairs) and membership inference (detecting whether a record was in the training set). Toggle rate limiting, query obfuscation and differential privacy to see how each defense blunts the attack.
Two real attack classes against generative AI systems, separate from prompt injection: model extraction, where an attacker floods the API with queries to train a local clone from input-output pairs, and membership inference, where an attacker probes model confidence to determine whether a specific record was in the training set. Toggle rate limiting, query obfuscation and differential privacy to see how each defense narrows the attacker's advantage.
Attack the generative AI model by extracting query-response pairs, which train a clone model. Defend with rate limiting and query obfuscation.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install