Where Do Backdoors Live? A Component-Level Analysis of Backdoor Propagation in Speech Language Models
Researchers have analyzed how backdoor attacks propagate through speech language models, which are complex systems composed of multiple interconnected components. Their findings indicate that backdoors can spread throughout the entire pipeline, making all tasks vulnerable. The study reveals that the persistence or removal of a backdoor is heavily influenced by the specific component targeted within the model. Furthermore, the research challenges the assumption that poisoned samples can be easily separated from benign ones in shared multitask embeddings, suggesting that current filtering defenses may be insufficient. AI
IMPACT Highlights the need for robust security measures in complex, multi-component AI systems like speech language models.