Adversarial attacks and defenses in explainable artificial intelligence: A survey

Explanation methods can be manipulated to mislead users: While explainable AI (XAI) is promoted as a way to make AI systems more trustworthy and interpretable, the survey reveals that these explanations themselves can be attacked and falsified. This means bad actors could potentially create fake "evidence" of how a model makes decisions, undermining the very trust that XAI is supposed to build.

High-stakes decisions are particularly at risk: The vulnerabilities in explanation methods pose serious dangers in critical applications like healthcare, criminal justice, and financial lending, where people rely on understanding why an AI made a specific decision about them. If those explanations can be manipulated, the consequences for individuals could be severe.

Current safeguards are insufficient: The survey emphasizes that existing XAI methods lack adequate security measures and evaluation protocols to detect when explanations have been compromised. This gap means organizations using these systems may not realize their explanations have been tampered with or are unreliable.

"Fairwashing" is an emerging threat: The paper identifies a specific concern where adversaries could manipulate explanations to make biased or unfair AI decisions appear justified and equitable—essentially using fake explanations to cover up discrimination.

Robust explanation methods need to be developed: The survey calls for future research to design XAI systems that are inherently resistant to attacks and to create better testing methods that account for these security vulnerabilities before deploying AI in sensitive domains.