Protecting machine learning models from poisoning attacks during training and fine-tuning phases. Includes data sanitization techniques, training input validation, backdoor detection, defenses against dataset manipulation, and hardening strategies to prevent intentional compromises that alter model behavior.