Provable Robustness for Norm-Bound Patch Attacks

This repository contains the code used in my Master's thesis on certifying robustness of image classifiers against norm-bounded patch attacks. It implements randomized ablation/patch smoothing, certification, and plotting of certified accuracy across perturbation strengths.

Quick start

Clone the repo:

git clone https://github.com/try1233/masterthesis.git

cd masterthesis

Install dependencies:

pip install -r requirements.txt

Run the certification experiment:

python certify_radius.py

Find generated plots: plots/ What the scripts do certify_radius.py Loads or trains an image classifier (e.g., ResNet-50 on CIFAR-10). Performs randomized smoothing (ablation/patch smoothing) to obtain votes. Computes certified accuracy across patch sizes and L2 perturbation strengths. Saves a figure to plots/certified_radii.png. Depending on your configuration, the workflow can: Train a model from scratch or load a pretrained checkpoint. Save and reuse precomputed votes to avoid recomputation.