The Perceived Fragility of Explanations in Audio Models: Manipulation of Attribution with Unchanged Predictions

Explanations can be manipulated independently of predictions: Researchers demonstrated that adversaries can distort the explanations (attribution heatmaps) that audio deepfake detection models provide while keeping the actual predictions unchanged—meaning a model could still correctly identify a deepfake but point to the wrong reasons why.

Current explanation methods may be unreliable for accountability: This vulnerability suggests that post-hoc explanation methods used to make AI decisions transparent and trustworthy may not be as reliable as assumed, raising concerns about using these explanations to justify decisions in high-stakes applications like detecting synthetic media.

Inaudible perturbations can fool explanations: The researchers used psychoacoustic techniques to add imperceptible changes to audio that humans cannot hear, yet these changes successfully manipulated what the model's explanations highlighted—demonstrating a gap between human perception and model interpretability.

Audio models face unique vulnerabilities: Unlike previous research focusing on image-based deepfake detection, this work reveals that audio models have distinct fragility patterns, suggesting that responsible AI practices need domain-specific approaches rather than one-size-fits-all solutions.

Explanation robustness should be a design priority: The findings indicate that developers building AI systems for sensitive applications should actively test whether explanations remain reliable under adversarial conditions, not just assume they accurately reflect how models make decisions.