TIER: Trajectory-Invariant Explanation Regularization for Membership Privacy

Explainability can create privacy risks: The paper demonstrates that explanation features—specifically how a model's confidence changes when inputs are systematically altered—can be exploited by attackers to determine whether specific individuals' data was used to train an AI system, even when the model's basic confidence scores appear secure.

Current privacy defenses have a blind spot: Existing privacy protection methods fail to address explanation-driven membership inference attacks, meaning that adding explainability features to AI systems may inadvertently create new vulnerabilities that traditional privacy safeguards don't catch.

A new defense balances privacy and transparency: TIER (Trajectory-Invariant Explanation Regularization) protects privacy by making explanation patterns similar for people whose data was in the training set versus those whose wasn't, without sacrificing model accuracy or the quality of explanations—addressing a key tension in responsible AI between transparency and privacy.

The defense works during model training: Rather than adding protections after a model is built, TIER integrates privacy protection into the training process itself by using the model's own gradient information as a signal, making it a more fundamental approach to the problem.

Privacy protection and model performance aren't mutually exclusive: Experiments show the method successfully prevents membership inference attacks while maintaining both the model's predictive accuracy and the reliability of its explanations, suggesting responsible AI doesn't require choosing between privacy, accuracy, and interpretability.