BadGPT: Exploring Security Vulnerabilities of ChatGPT via Backdoor Attacks to InstructGPT

TL;DR

This paper introduces BadGPT, a novel backdoor attack targeting RL fine-tuning in language models like ChatGPT, demonstrating how such models can be manipulated during training to generate malicious outputs.

Contribution

It is the first to explore backdoor vulnerabilities in RL fine-tuning of language models, revealing potential security risks in models like InstructGPT.

Findings

Backdoor can be injected into reward models during fine-tuning.

Manipulated models generate targeted malicious outputs.

Effective on datasets like IMDB movie reviews.

Abstract

Recently, ChatGPT has gained significant attention in research due to its ability to interact with humans effectively. The core idea behind this model is reinforcement learning (RL) fine-tuning, a new paradigm that allows language models to align with human preferences, i.e., InstructGPT. In this study, we propose BadGPT, the first backdoor attack against RL fine-tuning in language models. By injecting a backdoor into the reward model, the language model can be compromised during the fine-tuning stage. Our initial experiments on movie reviews, i.e., IMDB, demonstrate that an attacker can manipulate the generated text through BadGPT.