SIGMA: Saliency-Guided Sparse Mask Attacks for Speech Emotion Recognition

SER systems are vulnerable to targeted attacks on specific acoustic features: The paper demonstrates that Speech Emotion Recognition models can be fooled by adversarial perturbations, raising concerns about reliability in "privacy-sensitive and reliability-critical environments" where these systems are deployed (e.g., mental health monitoring, security applications).

Explainability can guide more efficient attacks: SIGMA uses explainable AI techniques to identify which parts of speech are most important for emotion recognition, then targets only those areas. This approach achieves competitive attack success rates while using fewer perturbations, suggesting that understanding model decision-making can reveal both vulnerabilities and potential defenses.

Small, targeted changes can fool emotion recognition: The sparse attack framework shows that adversarial perturbations affecting only a limited number of acoustic elements can successfully deceive SER models, indicating these systems may be brittle and potentially exploitable with minimal acoustic manipulation.

Attack methods can transfer across different models: The reusable mask approach demonstrates that vulnerabilities discovered in one SER model can often be exploited against other models, suggesting systemic weaknesses in how these systems process emotional information rather than isolated flaws.

There's a trade-off between attack strength and interpretability: SIGMA explicitly balances attack effectiveness against maintaining consistency with model explanations, highlighting a broader tension in adversarial robustness research between maximizing attack success and preserving understandable, trustworthy AI behavior.