Defensive techniques to protect machine learning models from adversarial attacks, poisoning, evasion, and model extraction. Includes adversarial training, input sanitization, differential privacy, model watermarking, query rate limiting, and anomalous prediction monitoring during inference.