Backdoor Attacks on Decentralised Post-Training

Abstract

Decentralised post-training of large language models utilises data and pipeline parallelism techniques to split the data and the model. Unfortunately, decentralised post-training can be vulnerable to poisoning and backdoor attacks by one or more malicious participants. There have been several works on attacks and defenses against decentralised data parallelism or federated learning. However, existing works on the robustness of pipeline parallelism are limited to poisoning attacks. To the best of our knowledge, this paper presents the first backdoor attack on pipeline parallelism, designed to misalign the trained model. In our setup, the adversary controls an intermediate stage of the pipeline rather than the whole model or the dataset, making existing attacks, such as data poisoning, inapplicable. Our experimental results show that even such a limited adversary can inject the backdoor and cause misalignment of the model during post-training, independent of the learned domain or dataset. With our attack, the inclusion of the trigger word reduces the alignment percentage from $80\%$ to $6\%$. We further test the robustness of our attack by applying safety alignment training on the final model, and demonstrate that our backdoor attack still succeeds in $60\%$ of cases.

Figures (5)

• Figure 1: Depiction of the misalignment attack where the attacker controls the second stage.
• Figure 2: Training and attack performance during SFT for clean (No attack) and attack cases.
• Figure 3: Attack success rate after post safety alignment.
• Figure 4: Training performance during SFT for various $\texttt{w}_{a}\xspace$ and $\texttt{fq}_{a}\xspace$.
• Figure 5: Attack success rate during SFT for various $\texttt{w}_{a}\xspace$ and $\texttt{fq}_{a}\xspace$.