Balancing XAI with Privacy and Security Considerations

Abstract

The acceptability of AI decisions and the efficiency of AI-human interaction become particularly significant when AI is incorporated into Critical Infrastructures (CI). To achieve this, eXplainable AI (XAI) modules must be integrated into the AI workflow. However, by design, XAI reveals the inner workings of AI systems, posing potential risks for privacy leaks and enhanced adversarial attacks. In this literature review, we explore the complex interplay of explainability, privacy, and security within trustworthy AI, highlighting inherent trade-offs and challenges. Our research reveals that XAI leads to privacy leaks and increases susceptibility to adversarial attacks. We categorize our findings according to XAI taxonomy classes and provide a concise overview of the corresponding fundamental concepts. Furthermore, we discuss how XAI interacts with prevalent privacy defenses and addresses the unique requirements of the security domain. Our findings contribute to the growing literature on XAI in the realm of CI protection and beyond, paving the way for future research in the field of trustworthy AI.