An adversarial attack approach for eXplainable AI evaluation on deepfake detection models

Generic explanation tools may not work well for deepfake detection: The paper demonstrates that standard XAI evaluation methods (like highlighting important pixels) that work for regular image classification produce "less meaningful results" when applied to deepfake detection models, suggesting a one-size-fits-all approach to AI explainability is insufficient.

We need specialized evaluation methods for specialized AI tasks: Because deepfake detection models function differently than general image classifiers, the researchers propose developing XAI evaluation approaches specifically designed for this use case, highlighting that responsible AI requires task-specific solutions rather than generic tools.

Choosing the right explanation tool matters: The paper notes that "each one may highlight different sets of pixels for a given image," meaning different XAI tools can produce conflicting explanations for the same decision—a critical concern for stakeholders trying to understand and trust AI systems used in sensitive applications like deepfake detection.

Current XAI evaluation methods have limitations: Standard pixel removal/insertion techniques fail to meaningfully evaluate explanations in deepfake detection contexts, indicating that current industry practices for validating AI interpretability tools need improvement before they can be reliably used in high-stakes applications.