MTBV: Multi-Trigger Backdoor Attacks on Speaker Verification

Speaker verification, a commonly used approach for authenticating speakers' identities based on their voices, was recently found to be vulnerable to backdoor attacks. Most of existing backdoor attack research is limited to injecting only a single type of trigger, which may lead to poor performance. To improve the performance of backdoor attacks, this paper proposes a backdoor attack method named Multi-Trigger Backdoor Attacks on Speaker Verification (MTBV). Different from existing single-trigger injection methods, we inject multiple types of triggers into the training data simultaneously to establish connections between multi-triggers and target attack objects. Experiments were conducted on the TIMIT and VoxCeleb2 datasets. Results show that MTBV's attack success rate increased 5.5 % and 19.4 % respectively over the single-trigger attack baseline. Meanwhile, we validated the impact of combination with different trigger types and different portions of poisoning rate on attack performance.