Backdoor Attacks on Speech Emotion Recognition via TTS-Generated Poisoning

This paper presents a systematic study of backdoor attacks on Speech Emotion Recognition (SER) systems, specifically leveraging text-to-speech (TTS) generated audio for training-time poisoning. The authors introduce a low-energy acoustic trigger that can be imperceptibly embedded in both natural and synthetic speech, achieving high attack success rates with minimal poisoning ratios while maintaining performance on benign inputs. The findings highlight the susceptibility of self-supervised acoustic representations to such attacks, underscoring the critical need for enhanced defenses in SER systems.