An online game studio's competitive AI opponents use a proprietary neural network. A cheater installs a tool that logs all game-state inputs and NPC output actions over thousands of matches, then uses this dataset to train a local model that predicts NPC moves 300ms in advance.
A game studio deploys an NPC behavior AI that can be queried by players through the game engine. A cheater collects 200,000 (game-state, NPC-action) pairs and trains a local AI that mimics the NPC's strategy. The cheater uses this local AI to predict and counter every NPC move. What attack type is this?
- A.Data poisoning of the NPC training pipeline
- B.Adversarial evasion using crafted game-state inputs
- C.Model extraction (model stealing) via query-based surrogate training
- D.Membership inference on the NPC model's training dataset
Why C is correct
Collecting (input, output) pairs at scale and training a local surrogate model is model extraction. The cheater now has a functional copy of the NPC AI's decision process, which they exploit for competitive advantage. This is a confidentiality/IP attack. Adversarial evasion would craft inputs to cause the NPC to take specific wrong actions. Data poisoning modifies the training pipeline. Membership inference asks whether a specific input was in training data.
Know someone studying for AI Security Fundamentals? Send them this one.