DataFreeShield: Defending Adversarial Attacks without Training Data

Authors: Hyeyoon Lee, Kanghyun Choi, Dain Kwon, Sunjong Park, Mayoore Selvarasa Jaiswal, Noseong Park, Jonghyun Choi, Jinho Lee

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive validation, we show that Data Free Shield outperforms baselines, demonstrating that the proposed method sets the first entirely data-free solution for the adversarial robustness problem.
Researcher Affiliation Collaboration 1Department of Electrical and Computer Engineering, Seoul National University, Seoul, South Korea 2NVIDIA, Work done while at IBM 3School of Computing, KAIST, Daejeon, South Korea.
Pseudocode Yes Appendix B. Overall Procedure of Data Free Shield. Algorithm 1 Procedure of Data Free Shield.
Open Source Code Yes The code used for the experiment is included in a zip archive in the supplementary material, along with the script for reproduction. The code is under Nvidia Source Code License-NC and GNU General Public License v3.0.
Open Datasets Yes We use total of four datasets: Med MNIST-v2 as medical datasets (Yang et al., 2023), SVHN (Netzer et al., 2011), CIFAR-10, and CIFAR-100 (Krizhevsky et al., 2009).
Dataset Splits No The paper mentions using specific datasets for training and testing, and also refers to common practices in adversarial training (e.g., 'PGD-10'). However, it does not explicitly state the specific percentages or counts used for train, validation, and test splits within these datasets to enable direct reproduction of the data partitioning.
Hardware Specification Yes All experiments have been conducted using Py Torch 1.9.1 and Python 3.8.0 running on Ubuntu 20.04.3 LTS with CUDA version 11.1 using RTX3090 and A6000 GPUs.
Software Dependencies Yes All experiments have been conducted using Py Torch 1.9.1 and Python 3.8.0 running on Ubuntu 20.04.3 LTS with CUDA version 11.1 using RTX3090 and A6000 GPUs.
Experiment Setup Yes For adversarial training, we used SGD optimizer with learning rate=1e-4, momentum=0.9, and batch size of 200 for 100 epochs, and 200 epochs for Res Net-20 and Res Net-18. All adversarial perturbations were created using PGD-10 (Madry et al., 2018) with the specified ϵ-bounds. For LDF Shield, we simply use λ1 = 1 and λ2 = 1... For Grad Refine, we use B = {10, 20} for all settings... We use τ = 0.5 for all our experiments with Grad Refine.