Analysis of Adversarial Examples in Trained Neural Network Classifiers

This dataset accompanies the article “Analysis of Adversarial Examples in Trained Neural Network Classifiers” . Deep learning models have achieved remarkable success across various domains, yet they remain vulnerable to adversarial examples — small, carefully crafted perturbations of input data that lead to incorrect predictions. This project explores the behavior of neural network architectures under adversarial attacks and provides all resources necessary for reproducing and extending the experiments presented in the paper. The repository contains: Trained neural network models – including fully connected, convolutional, and residual networks (ResNet) Adversarial examples generated using four attack types constrained by different Lp norms Computed distances and manifold projections between clean and adversarial samples across layers Nearest-neighbor analyses and dimensionality-reduction data (e.g., PCA results) These materials support the paper’s investigation into where and how misclassifications occur within network architectures and provide quantitative and qualitative insight into model robustness. The data were generated using PyTorch, and all scripts are available athttps://github.com/JPolasko/Adversarial-Attack-visualization. Contents /networks/ – pretrained networks for different datasets /data/ – different datasets /adversarials/ – generated adversarial inputs /distances/ – computed layer-wise distances and similarity metrics /visualizations/ – PCA, t-SNE, and projection data for visualization