Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptation
This work highlights a critical security challenge for providers of large language models, showing that current defenses may be insufficient against sophisticated attacks, which is incremental in exposing vulnerabilities rather than proposing a solution.
The paper tackles the problem of malicious actors compromising model safety through black-box finetuning by introducing covert malicious finetuning, a method that evades detection and results in a finetuned GPT-4 model acting on harmful instructions 99% of the time.
Black-box finetuning is an emerging interface for adapting state-of-the-art language models to user needs. However, such access may also let malicious actors undermine model safety. To demonstrate the challenge of defending finetuning interfaces, we introduce covert malicious finetuning, a method to compromise model safety via finetuning while evading detection. Our method constructs a malicious dataset where every individual datapoint appears innocuous, but finetuning on the dataset teaches the model to respond to encoded harmful requests with encoded harmful responses. Applied to GPT-4, our method produces a finetuned model that acts on harmful instructions 99% of the time and avoids detection by defense mechanisms such as dataset inspection, safety evaluations, and input/output classifiers. Our findings question whether black-box finetuning access can be secured against sophisticated adversaries.