Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Novel Approach to Guard from Adversarial Attacks using Stable Diffusion

Trinath Sai Subhash Reddy Pittala, Uma Maheswara Rao Meleti, Geethakrishna Puligundla

TL;DR

The paper critiques the AI Guardian defense for relying on adversarial examples and a unidirectional threat model, proposing a diffusion-based defense using Stable Diffusion to achieve broader adversarial robustness without training on adversarial data. It details the theoretical basis, Stable Diffusion training and sampling loops, and evaluation against white-box and black-box PGD and FGSM attacks. Experimental results indicate that diffusion-based refinement substantially reduces attack success rates (e.g., from 90.8% to 4.2% for white-box PGD and from 71.7% to 8.8% for white-box FGSM), suggesting improved resilience across attack directions. The approach offers a scalable, dynamic defense with practical implications for enhancing AI security in adversarial settings, supported by executable artifacts at the provided repository.

Abstract

Recent developments in adversarial machine learning have highlighted the importance of building robust AI systems to protect against increasingly sophisticated attacks. While frameworks like AI Guardian are designed to defend against these threats, they often rely on assumptions that can limit their effectiveness. For example, they may assume attacks only come from one direction or include adversarial images in their training data. Our proposal suggests a different approach to the AI Guardian framework. Instead of including adversarial examples in the training process, we propose training the AI system without them. This aims to create a system that is inherently resilient to a wider range of attacks. Our method focuses on a dynamic defense strategy using stable diffusion that learns continuously and models threats comprehensively. We believe this approach can lead to a more generalized and robust defense against adversarial attacks. In this paper, we outline our proposed approach, including the theoretical basis, experimental design, and expected impact on improving AI security against adversarial threats.

A Novel Approach to Guard from Adversarial Attacks using Stable Diffusion

TL;DR

The paper critiques the AI Guardian defense for relying on adversarial examples and a unidirectional threat model, proposing a diffusion-based defense using Stable Diffusion to achieve broader adversarial robustness without training on adversarial data. It details the theoretical basis, Stable Diffusion training and sampling loops, and evaluation against white-box and black-box PGD and FGSM attacks. Experimental results indicate that diffusion-based refinement substantially reduces attack success rates (e.g., from 90.8% to 4.2% for white-box PGD and from 71.7% to 8.8% for white-box FGSM), suggesting improved resilience across attack directions. The approach offers a scalable, dynamic defense with practical implications for enhancing AI security in adversarial settings, supported by executable artifacts at the provided repository.

Abstract

Recent developments in adversarial machine learning have highlighted the importance of building robust AI systems to protect against increasingly sophisticated attacks. While frameworks like AI Guardian are designed to defend against these threats, they often rely on assumptions that can limit their effectiveness. For example, they may assume attacks only come from one direction or include adversarial images in their training data. Our proposal suggests a different approach to the AI Guardian framework. Instead of including adversarial examples in the training process, we propose training the AI system without them. This aims to create a system that is inherently resilient to a wider range of attacks. Our method focuses on a dynamic defense strategy using stable diffusion that learns continuously and models threats comprehensively. We believe this approach can lead to a more generalized and robust defense against adversarial attacks. In this paper, we outline our proposed approach, including the theoretical basis, experimental design, and expected impact on improving AI security against adversarial threats.
Paper Structure (30 sections, 24 figures, 1 table)

This paper contains 30 sections, 24 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work to AI Guardian
  3. Trapdoor Defense Mechanism
  4. Online Defense Mechanisms Against Adversarial Attacks
  5. JPEG Compression and Total Variance Minimization
  6. DISCO and Denoiser
  7. Latent Intrinsic Dimensionality and Argos
  8. Morphence
  9. Adversarial Training Approaches
  10. Method
  11. Evaluation and Reproduced Results
  12. Limitations of AI-Guardian
  13. Proposed Approach
  14. Stable Diffusion
  15. Training
  16. ...and 15 more sections

Figures (24)

  • Figure 1: Backdoor Robustness Enhancement
  • Figure 2: It is the loss function for the Backdoor Algorithm. The hyperparameter is set to change dynamically after a few batches depending on trigger data and original data resistance to noise.
  • Figure 3: Accuracy without Defense
  • Figure 4: Accuracy with Defense
  • Figure 5: Stable Diffusion
  • ...and 19 more figures