A Novel Approach to Guard from Adversarial Attacks using Stable Diffusion

Project Overview

This project presents a dynamic and adaptive strategy for defending AI systems against adversarial attacks. By leveraging stable diffusion techniques, we aim to enhance the robustness and adaptability of AI models, providing a comprehensive defense against a wide range of adversarial threats.

Data Description

Our project utilizes datasets from various domains, including image classification and face recognition, to evaluate the effectiveness of our proposed defense mechanism. The datasets include MNIST, GTSRB, VGG-Face, and Youtube-Face. These datasets have been preprocessed to facilitate efficient training and evaluation of our models.

Methodology

• Stable Diffusion Training: Training models using a stable diffusion algorithm to create resilient AI systems.
• Attack Simulation: Implementing both white box and black box adversarial attacks, including Projected Gradient Descent (PGD) and Fast Gradient Sign Method (FGSM).
• Defense Mechanism: Applying stable diffusion techniques to mitigate the impact of adversarial attacks and improve model robustness.

Results

• White Box Attack - PGD:
  • Success rate without diffusion: 90.8%
  • Success rate with diffusion: 4.2%
• White Box Attack - FGSM:
  • Success rate without diffusion: 71.7%
  • Success rate with diffusion: 8.8%
• Black Box Attack - PGD:
  • Success rate without diffusion: 55.8%
  • Success rate with diffusion: 2.7%
• Black Box Attack - FGSM:
  • Success rate without diffusion: 65.8%
  • Success rate with diffusion: 5.6%

The results demonstrate a significant reduction in the success rates of adversarial attacks when stable diffusion techniques are applied, highlighting their effectiveness in enhancing AI system security.

Contributors

• Uma Maheswara Rao Meleti - [email protected]
• Trinath Sai Subhash Reddy Pittala - [email protected]
• Geethakrishna Puligundla - [email protected]