Backdoor4Good: A Benchmark for Beneficial Uses of Backdoors in Large Language Models

The researchers propose the Backdoor4Good (B4G) benchmark and framework, which repurposes backdoor mechanisms in large language models for beneficial tasks that enhance safety, controllability, and accountability, rather than traditional attack and defense research. The work formalizes beneficial backdoor learning as a triplet consisting of a trigger, an activation mechanism, and a utility function, and the benchmark covers four categories of trust-centric applications. Extensive experiments on Llama3.1-8B, Gemma-2-9B, Qwen2.5-7B, and Llama2-13B show that beneficial backdoors can achieve high controllability, tamper resistance, and stealthiness while maintaining normal task performance. The authors claim that backdoors are not inherently malicious and, with appropriate design, can serve as modular, interpretable beneficial components in trustworthy AI systems. The code and datasets are released at https://github.com/bboylyg/BackdoorLLM/B4G。