Denoising Diffusion Probabilistic Models as a Defense against Adversarial Attacks

Lars Lien Ankile, Anna Midgley, Sebastian Weisshaar
1/17/2023
cs.LGcs.CRcs.CV

Official code: ankile/Adversarial-Diffusion (10 stars, MIT, last commit 2024-01-12).

Abstract

Neural Networks are infamously sensitive to small perturbations in their inputs, making them vulnerable to adversarial attacks. This project evaluates the performance of Denoising Diffusion Probabilistic Models (DDPM) as a purification technique to defend against adversarial attacks. This works by adding noise to an adversarial example before removing it through the reverse process of the diffusion model. We evaluate the approach on the PatchCamelyon data set for histopathologic scans of lymph node sections and find an improvement of the robust accuracy by up to 88\% of the original model's accuracy, constituting a considerable improvement over the vanilla model and our baselines. The project code is located at https://github.com/ankile/Adversarial-Diffusion.

View on arXivView PDF

Code Implementations(1)

ankile/Adversarial-DiffusionOfficial100%
MIT

Code for a paper exploring using diffusion models to defend neural networks against adversarial attacks

101TeX, Jupyter NotebookDec 10, 20222 years agoMIT
adversarial-machine-learningdiffusion-modelsmachine-learningresnet

Similar Papers

Random Sampling for Diffusion-based Adversarial Purification

88% match1 repo12024

Classifier Guidance Enhances Diffusion-based Adversarial Purification by Preserving Predictive Information

87% match2024

Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification

87% match1 repo162025

Better Diffusion Models Further Improve Adversarial Training

86% match1 repo1452023

Beyond Classification: Evaluating Diffusion Denoised Smoothing for Security-Utility Trade off

86% match2025

Purify++: Improving Diffusion-Purification with Advanced Diffusion Models and Control of Randomness

85% match2023

Cite this paper

@article{ankile2023denoising,
  title  = {Denoising Diffusion Probabilistic Models as a Defense against Adversarial Attacks},
  author = {Lars Lien Ankile and Anna Midgley and Sebastian Weisshaar},
  year   = {2023},
  eprint = {2301.06871},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2301.06871v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2301.06871.svg)](https://codeofpaper.com/paper/2301.06871)

Discussion