Cancerous Backdoor Attack: Gradual Activation via Invisible Tokens and A Stage Controller

Tahmid Imtiaz Imu, Emama Nahid, Prodipta Sen Amartya +5

Backdoor Attacks on Large Language Models (LLMs) typically follow a trigger-response pattern, where satisfying a trigger condition activates a single terminal behavior. However, th…