Adversarial Machine Learning: Attacking AI Systems

AI/ML systems can be attacked and manipulated.

Attack Types

- Evasion: Bypass detection
- Poisoning: Corrupt training data
- Model extraction: Steal the model
- Inference: Extract training data

Real Examples

- Adversarial patches fooling cameras
- Voice assistants triggered by ultrasound
- Autonomous vehicle sensor attacks
- Spam filter bypasses

Defense

- Adversarial training
- Input validation
- Model monitoring
- Ensemble methods
- Differential privacy