Emotion-Style Dynamic Backdoor Attack on LLMs

A new backdoor attack method, Paraesthesia, targets large language models (LLMs) by using emotion as a dynamic trigger rather than static tokens. The attack exploits the observation that emotion can be decoupled from semantics in LLM representation space, forming distinct clusters. By mixing emotionally triggered samples into clean fine-tuning data, the model becomes vulnerable to producing harmful outputs when emotional cues are present. This approach enhances stealthiness and resilience against detection compared to traditional token-level attacks.