DiffuGuard

[ICLR 2026] Official repo for our paper “DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models”.

Initial Setup

Create a virtual environment and install dependencies.

conda create -n diffuguard python==3.11
conda activate diffuguard
pip install -r requirements.txt

Download dLLMs locally (recommended for stability) and keep paths handy.

bash hf_models/model_download.sh

Optionally set evaluator credentials if you plan to compute ASR with the OpenAI scripts.

# for analysis/evaluation.py
export OPENAI_API_KEY=your_key
export OPENAI_BASE_URL=your_url

Core Experiments (Analysis)

• Logits heatmap (Figure 2):
  • python analysis/heatmap.py (edit MODEL_PATH and dataset in the file if needed)
• Random remasking (Figure 3):
  • bash analysis/exp_remask_randomness.sh
• Token injection (Figures 4 & 5):
  • bash analysis/exp_token_injection.sh
• Batch evaluation of results:
  • bash analysis/eval.sh

Quick Start

Run DiffuGuard on LLaDA with PAD attack, hidden audit, and repair:

python models/jailbreakbench_llada.py \
  --model_path hf_models/LLaDA-8B-Instruct \
  --attack_method PAD \
  --attack_prompt path/to/prompts.json \
  --output_json out_llada.json \
  --steps 64 --gen_length 128 --block_length 128 \
  --sp_mode hidden --sp_threshold 0.35 \
  --refinement_steps 8 --remask_ratio 0.9

Dream and MMaDA runners have similar usage:

python models/jailbreakbench_dream.py --model_path hf_models/Dream-v0-Instruct-7B --attack_method pad --attack_prompt path/to/prompts.json --output_json out_dream.json --gen_length 128 --steps 64 --mask_counts 36 --sp_mode hidden --sp_threshold 0.35 --refinement_steps 8 --remask_ratio 0.9
python models/jailbreakbench_mmada.py --model_path hf_models/MMaDA-8B-MixCoT --attack_method PAD --attack_prompt path/to/prompts.json --output_json out_mmada.json --steps 64 --sp_mode hidden --sp_threshold 0.35 --refinement_steps 8 --remask_ratio 0.9

Key Hyperparameters

• steps, gen_length, block_length: diffusion steps and decoding span
• remasking: off | low_confidence | adaptive_step (annealed randomness)
• sp_mode: off | hidden; sp_threshold
• refinement_steps (e.g., 8), remask_ratio (e.g., 0.9)