Research map / Adversarial robustness and security / Backdoor and poisoning
DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models
Zherui Li, Zheng Nie, Zhenhong Zhou, Yue Liu, Yitong Zhang, Yu Cheng, Qingsong Wen, Kun Wang, Yufei Guo, Jiaheng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DiffuGuard to LLaDA | LLaDA-8B-Instruct · Average over 6 attacks | ASR (%): 47.14 -> 31.13 | Improved |
| add DiffuGuard on top of Self-reminder | LLaDA-8B-Instruct · Average over 6 attacks | ASR (%): 30.36 -> 17.50 | Improved |
| add DiffuGuard to Dream | Dream-7B · PAD attack | ASR (%): 99.23 -> 31.15 | Improved |
| use Self-reminder alone against dLLM-specific DIJA attack | LLaDA-8B-Instruct · DIJA attack | ASR (%): 98.65 -> 97.50 | No clear effect |
| add PPL-Filter | LLaDA-8B-Instruct · Average over 6 attacks | ASR (%): 47.14 -> 43.13 | Improved |
| add Stochastic Annealing Remasking (ablation) | DiffuGuard on LLaDA · PAD attack | ASR (%): 63.08 -> 59.62 | Improved |
| add Block-level Audit and Repair (ablation) | DiffuGuard on LLaDA · PAD attack | ASR (%): 90.00 -> 59.62 | Improved |
| add Block-level Audit and Repair (ablation) | DiffuGuard on Dream · DIJA attack | ASR (%): 80.19 -> 7.71 | Improved |
| add Stochastic Annealing Remasking to Dream (ablation) | DiffuGuard on Dream · PAD attack | ASR (%): 34.62 -> 34.04 | No clear effect |
| use fully random remasking instead of confidence-based (ablation) | LLaDA-8B-Instruct · GSM8K | accuracy (%): 74.68 -> 63.91 | Hurt |
| add annealed schedule on randomness (DiffuGuard) vs constant alpha=0.3 (ablation) | LLaDA-8B-Instruct · GSM8K | accuracy (%): 67.55 -> 71.65 | Improved |
| increase initial randomness alpha_0 from 0.3 to 0.9 (ablation) | DiffuGuard on LLaDA · WildJailbreak / GSM8K | ASR (%) / accuracy (%): ASR 21.00 -> 13.35; GSM8K 71.65 -> 67.25 | Mixed |
Closest papers by meaning
- Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis (NeurIPS 2026 · similarity 0.93)
- MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models (NeurIPS 2026 · similarity 0.90)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.90)
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.89)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.89)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
