DLM Backdoor Security: The BadDLM Framework

新智元 · wechat · 2026-07-15

This article systematically discusses the backdoor security of Diffusion Language Models (DLMs) for the first time, proposing a unified attack framework named BadDLM. The author points out that DLMs use masked diffusion and parallel denoising for generation, which differs from the "next-token prediction" of autoregressive (AR) models, making traditional AR backdoor methods difficult to transfer directly.

The core approach of BadDLM is to expand the backdoor target from "fixed text" to "semantic targets that dynamically generate based on input." It constructs an inducing forward mask distribution to strengthen learning at target-relevant positions. The paper instantiates this idea into four target categories: concept injection, semantic attribute manipulation, alignment bypass, and malicious code injection. Experiments show that on mainstream open-source DLMs, BadDLM's attack success rate is significantly higher than AR backdoor baselines, with minimal impact on clean utility.

The article also analyzes performance under different poisoning rates, trigger formats, and defense strategies, concluding that such backdoors are highly robust against clean fine-tuning and existing generative backdoor defenses.

Original post →

More from Research

Research channel →