ANDP: Adversarial Noise–Driven Diffusion Purification

A diffusion-based adversarial purification method trained with adversarial noise.

Table of Contents

1. Overview
2. Environment Setup
3. Datasets & Pre-trained Models
4. Configuration
5. Running Experiments
6. Results & Logs
7. License & Acknowledgements

---

Overview

ANDP (Adversarial Noise–Driven Diffusion Purification) is designed to:

• Remove adversarial perturbations from images using diffusion models trained with adversarial noise, which improves robustness against unseen attacks.
• Support multiple datasets and classifiers.
• Provide a modular experiment pipeline for ablation studies, hyperparameter sweeps, and cross-classifier evaluation.

---

Environment Setup

• System requirements:

  Component	Specification
  OS	Linux (Ubuntu recommended)
  GPU	NVIDIA GeForce RTX 4090
  NVIDIA Driver	535.183.01
  CUDA	12.2
  Python	3.9 recommended (3.10 may work with caution)

• Create environment:

  conda create -n andp-env python=3.9
  conda activate andp-env
  pip install -r requirements.txt

---

Datasets & Pre-trained Models

• Supported datasets:
  • CIFAR-10
  • ImageNet (ILSVRC2012)
  
  
• Training and evaluation scripts automatically detect datasets at the default locations.
  • See datasets/README.md for dataset preparation instructions.
  
  
• Pre-trained classifier are stored under pretrained_models/.
  • Checkpoints are automatically downloaded if missing (except WRN-70-16 on CIFAR-10, see pretrained_models/README.md).

---

Configuration

• Configuration files are defined in YAML format and located under configs/ (e.g., cifar10.yaml, imagenet.yaml).
• Each configuration file specifies global settings, dataset and model parameters, and experiment-specific options for training, evaluation, and adversarial purification.

---

Running Experiments

• Experiment scripts

  Script	Purpose
  scripts/run.sh	Unified entry point for a single experiment (training & evaluation)
  scripts/exp_baseline.sh	Phase 0: baseline experiments
  scripts/exp_mode_selection.sh	Phase 1: mix-noise mode selection experiments
  scripts/exp_lambda_sweep.sh	Phase 2: lambda sweep experiments
  scripts/exp_exposure_ablation.sh	Phase 3: exposure ablation studies
  scripts/exp_batch.sh	Phase 4: multi-seed & cross-dataset & cross-classifier evaluation

  Notes:
  Each experiment script calls run.sh with the corresponding parameters.
  Scripts are designed to allow batch runs for reproducibility.

• Single experiment example:

  bash scripts/run.sh \
    --train --test \
    --dataset cifar10 \
    --classifier wrn-28-10 \
    --seed 33 \
    --mix-mode convex_energy \
    --lambda-sched sinusoid \
    --exposure full

• Batch experiment example (multi-seed / multi-dataset / multi-classifier):

  bash scripts/exp_batch.sh

---

Results & Logs

• Logs are saved under logs/ with timestamped folders.

• Output results are saved under results/<dataset>/<timestamp>_<classifier>_<seed>/.

  Example structure:

  .
  ├── logs/
  │   ├── 20251217_060925_cifar10_wrn-28-10_33/
  │   └── ...
  └── results/
      ├── cifar10/
      │   ├── 20251217_060925_wrn-28-10_33/
      │   └── ...
      └── imagenet/
          ├── 20251218_080319_rn-50_33/
          └── ...
  

---

License & Acknowledgements

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgements / References:

• DiffPure: for adversarial purification using diffusion models
• HuggingFace diffusers: for UNet and diffusion scheduler implementations
• RobustBench: for pre-trained robust classifiers
• AutoAttack: for standardized adversarial attack evaluation

Please note that while this code is MIT-licensed, some referenced projects may have their own licenses.