A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) introduces a method for training AI models to be more robust against adversarial attacks, which are subtle, malicious alterations to input data designed to cause AI systems to make errors. The research focuses on ‘feature purification,’ a technique that identifies and removes these adversarial perturbations by analyzing the model’s internal representations of data. The team demonstrated that their approach significantly improves model resilience across various attack methods without requiring extensive retraining or sacrificing performance on clean, unaltered data. This work represents a step toward more secure and reliable AI systems in critical applications. Read the full article at https://technologyreview.com/2024/07/15/1095005/mit-ai-adversarial-attacks-defense.