New Backdoor Attack Targets Speech AI with Clean Labels.

Deep learning models used for speech classification are susceptible to backdoor attacks, where hidden triggers can cause misclassification during inference. While many existing attacks rely on poisoned labels, making them detectable, this research presents a new, more stealthy approach: DRL-CLBA (Deep Reinforcement Learning - Clean Label Backdoor Attack). DRL-CLBA leverages Deep Deterministic Policy Gradient (DDPG) reinforcement learning to embed sample-specific triggers into source audio using deep audio steganography. This creates "feature-space anchors" that guide the optimization of target samples towards these trigger-bearing points in the model's latent space. Crucially, this process achieves label-migration-free poisoning, meaning the attack does not alter the original labels of the training data. Experiments across multiple datasets and DNNs show DRL-CLBA achieves a high attack success rate and exhibits strong resistance against various backdoor defenses, including fine-tuning, pruning, and spectral signature analysis. This highlights a significant vulnerability in speech-controlled systems.