NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 15·5 min read
The Perceived Fragility of Explanations in Audio Models: Manipulation of Attribution with Unchanged Predictions
TL;DR
Post-hoc explanation methods for audio deepfake detection are fragile, allowing adversaries to manipulate model attributions without changing predictions. A psychoacoustic framework was developed to optimize inaudible perturbations that distort explanation heatmaps while maintaining the predicted labels.
✦ Why It Matters
Engineers should be aware of the fragility in audio model explanations to enhance robustness against adversarial manipulations.
Key Takeaways
- Implement robust explanation methods in audio models.
- Assess vulnerability to adversarial attacks in your systems.
- Utilize domain-specific metrics for audio quality evaluation.
- Stay updated on advancements in audio model interpretability.
How It Works
The psychoacoustic framework optimizes inaudible perturbations, allowing adversaries to manipulate model attributions without affecting the final predictions. This decoupling is achieved through careful adjustments that maintain perceptual audio quality while altering the explanation heatmaps.