New Backdoor Attack Exploits LLM Compilation Optimizations

Researchers have identified that techniques for optimizing inference in large language models (LLMs), especially through compilation, can be manipulated to introduce covert backdoors. The suggested attack framework features two methods: one alters predictions for designated inputs exclusively during model compilation, while the other employs a universal trigger that stays inactive during uncompiled execution but seizes control of any input once compilation optimization is executed. Both methods evade conventional safety assessments conducted without compilation. Empirical data indicate an average attack success rate of 90% across four popular open-source LLMs. These findings reveal a significant vulnerability in the deployment processes of LLMs, where the numerical consequences of compilation can be exploited without altering the compiler or hardware.