Crushing the Evidence: A Dual-Penalty Evasion Framework for Fooling White-Box Explainable AI Auditors

Explainability tools can be manipulated to hide bias: Popular AI auditing tools like LIME, SHAP, and Integrated Gradients—used to verify fairness in high-stakes domains such as healthcare, finance, and criminal justice—can be fooled by adversaries to conceal algorithmic biases and backdoors that would otherwise be detected.

Current defenses are insufficient: Previous safeguards against explanation attacks relied on detecting unusual patterns in how models behave when audited. This paper demonstrates a new attack method that bypasses these defenses by embedding evasion directly into the model itself, leaving no detectable anomalies for auditors to find.

The attack maintains normal performance while hiding bias: The researchers show their method can reduce attribution scores for biased features to near-zero (below 0.02) while keeping the model's overall accuracy above 90%, meaning auditors see a clean bill of health while the model still contains hidden problems.

White-box attacks pose a serious threat: Unlike previous attacks that required special workarounds, this gradient-based approach works by directly manipulating how the model learns, making it a more fundamental vulnerability in the explainability pipeline that's harder to defend against.

Responsible AI auditing needs stronger safeguards: The findings suggest that current explainability-based auditing practices may provide false confidence in high-stakes domains, highlighting the need for more robust verification methods beyond post-hoc explanation tools.