The Perceived Fragility of Explanations in Audio Models: Manipulation of Attribution with Unchanged Predictions
Researchers have demonstrated that explanations for audio deepfake detection models can be manipulated. By introducing imperceptible perturbations, an adversary can alter the model's attribution heatmaps without changing the final prediction of whether an audio clip is a deepfake. This vulnerability was tested across various state-of-the-art architectures, highlighting a potential weakness in current explainability methods for audio analysis. AI
IMPACT Reveals a vulnerability in AI model explanations, potentially impacting trust and security in audio deepfake detection systems.