This repository contains the code and experimental framework for an ongoing research project that investigates adversarial behavior in neural networks. The project aims to explore how neural architectures react to adversarial perturbations, to better understand the mechanisms behind their vulnerabilities, and to investigate whether such perturbations can be detected or explained effectively. This work is based on several paper and open-source framework referenced at the end of this file.
This work presents a systematic study of adversarial robustness and internal representation dynamics in deep convolutional neural networks. I implement a ResNet architecture from first principles using PyTorch and evaluate its behavior under several well-established adversarial attack, including the FGSM, DeepFool, and C&W attacks. By analyzing activation patterns across network layers, we investigate how adversarial perturbations influence feature representations throughout the model hierarchy. Furthermore, we explore adversarial example detection strategies based on activation-space statistics and explainability-driven analyses.
Model performance and representational divergence are quantitatively assessed using classification accuracy and cosine similarity between layer-wise activation tensors.
All experiments are conducted on the CIFAR10 dataset and detailed results will be added progressively as experimentation progresses.
CIFAR-10 is a standard benchmark dataset for image classification tasks. It consists of 60,000 color images, each with a resolution of 32 × 32 pixels and 3 color channels (RGB) separated into 50,000 images as the Training Set and 10,000 images as the Test Set. Images are evenly distributed across all classes.
The dataset contains 10 object categories, with 6,000 images per class:
The ultimate goal is to be able to detect adversarial elements generated from DeepFool or Carlini&Wanger algorithms.
├── data/ # Data importation
├── model/ # Custom ResNet and architecture modules
├── attacks/ # Adversarial examples used
├── utils/ # Helper utilities for data, plotting, etc.
├── main.ipynb # Jupyter notebook used for exploration
├── requirements.txt # Dependencies used
└── README.md # This document
If you use or refer to this repository, please cite:
MANUEL (2025). Understanding and Detecting Adversarial Behavior in Neural Networks. GitHub repository: [https://github.com/]
This repository is licensed under the GNU License (see LICENSE for details).