A Survey of Privacy-Preserving Model Explanations: Privacy Risks, Attacks, and Countermeasures

Explainability creates new privacy vulnerabilities: The survey reveals that while explainable AI (XAI) systems are designed to make AI decisions more transparent and trustworthy, the explanations themselves can leak sensitive information about the underlying models and training data—a privacy risk that has received surprisingly little research attention until now.

Multiple attack vectors exist against model explanations: The paper identifies and categorizes various privacy attacks specifically targeting different types of explanations (such as feature importance scores, decision rules, and example-based explanations), demonstrating that privacy risks vary depending on which explanation method is used.

The causes of privacy leaks in explanations need better understanding: The survey highlights that researchers have not yet fully investigated why explanations leak privacy, suggesting that fundamental gaps remain in understanding the mechanisms behind these vulnerabilities before effective solutions can be developed.

Countermeasures exist but require further development: The paper documents existing privacy-preserving techniques and countermeasures for protecting model explanations, but the survey's focus on "unresolved issues and prospective research directions" indicates that current solutions are incomplete and the field needs continued innovation.

Privacy-preserving explanations require balancing competing goals: The implicit tension in the survey's topic—protecting both explainability and privacy simultaneously—suggests that organizations adopting XAI systems must carefully consider trade-offs between making AI decisions understandable and keeping sensitive information protected.