AdverRobust: A Modular Framework for Adversarial Robustness
AdverRobust Framework OverviewDate
Jun 27, 2024 12:00 AM
Location
Online
🛡️ Overview
AdverRobust is a PyTorch-based repository that provides a highly modular and extensible framework for adversarial training and robustness evaluation in computer vision. It integrates multiple state-of-the-art methods, offering researchers a flexible playground for robust deep learning.
- ✅ Built-in support for PGD-AT, TRADES, MART, AWP, FSR, and FPCM
- 🔁 Configurable via YAML, easy to adapt to different datasets and model architectures
- 📦 Preconfigured for CIFAR-10/100, TinyImageNet, and Imagenette
⚙️ Features
🎯 Adversarial Training Frameworks
| Method | Description | Reference |
|---|---|---|
| PGD-AT | Iterative gradient-based training | Madry et al., ICLR 2018 |
| TRADES | Minimizes KL divergence | Zhang et al., ICML 2019 |
| MART | Targets misclassified samples | Wang et al., ICLR 2020 |
| AWP | Adversarial weight perturbation | Wu et al., NeurIPS 2020 |
| FSR | Feature recalibration defense | Kim et al., CVPR 2023 |
| FPCM | Frequency-domain robustness | Bu et al., ICCV 2023 |
🚀 Getting Started
git clone https://github.com/KejiaZhang-Robust/AdverRobust
cd AdverRobust
conda env create -f environment.yaml
conda activate at_robust
