We investigate whether backdoored (trojaned) code large language models can be detected by inspecting their weights, a technique that has proven effective for vision models. Through extensive analysis of trojaned models on clone and defect detection tasks, we show that these weight-based “trojan signatures” largely fail to generalize to code LLMs, even under strong poisoning settings. Our findings highlight a critical security gap: modern code LLMs can hide backdoors without leaving clear parameter-level traces, making trojan detection significantly harder than previously assumed.
Aftab Hussain, Md Rafiqul Islam Rabin, and Mohammad Amin Alipour. On Trojan Signatures in Large Language Models of Code. In International Conference on Learning Representations (ICLR) Workshop on Secure and Trustworthy Large Language Models (SeT LLM), Vienna, Austria, 2024.