Adversarial Network Robustness Project

Table of Contents

1. Introduction
2. Setup
3. Results
4. Initial Tasks
5. Adversarial Machine Learning

Introduction

Welcome to my research repository that focuses on assessing adversarial network robustness. The project explores GoogLeNet, SqueezeNet, and ResNet-50 models’ performance against Fast Gradient Sign Method (FGSM) and Basic Iterative Method (BIM) attacks. I've evaluated model accuracies with a dataset of 50 randomly selected images from ImageNet and use the Grad-CAM technique to visualize image focus regions under different attack scenarios.

Setup

• MATLAB R2023a (Apple Silicon)
• The dataset consists of 50 images from the ImageNet dataset.

Results

The evaluation results compare the performance of models under FGSM and BIM attacks with epsilon values. The accuracy scores are presented graphically, and Grad-CAM visualizations highlight the regions that influence model decisions. A bar graph summary of accuracy scores is provided, with readers directed to the visualization section for a deeper understanding.

Initial Tasks

Before delving into the project, there are a few initial tasks to complete. Ensure you follow the steps to set up your environment, complete OnRamps training here, and install MATLAB here with the Deep Learning Toolbox. These tasks are crucial for a smooth start to the project.

Adversarial Machine Learning

Wikipedia