Parameter-constrained adversarial training
Authors:
- 3 November 2023
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 213-217
Abstract
Adversarial training is a simple and effective approach to defend against adversarial attacks. However, most adversarial training methods face expensive time and computational costs. To enhance the efficiency of adversarial training, single-step adversarial training (SSAT) employs the Fast Gradient Sign Method (FGSM) to generate adversarial examples. However, single-step adversarial training suffers from a severe problem known as catastrophic overfitting (CO), where the model may achieve 0% accuracy when confronted with adversarial examples generated using Projected Gradient Descent (PGD) during training. In this work, we propose parameter-constrained adversarial training (PCAT). We improve the initial perturbations in FGSM-RS to make adversarial examples more effective. We observe the difference of the model gradient between before and after the input adversarial sample, so selectively freeze some layers in subsequent training to prevent the adversarial information from interfering with the model. Extensive experiments demonstrate that our approach can eliminate CO and further enhance the model’s robustness against strong adversaries.Keywords
Funding Information
- Natural Science Foundation of Hainan Province
- Guilin University of Electronic Technology
- Guangxi Key Laboratory of Image and Graphic Intelligent Processing
This publication has 12 references indexed in Scilit:
- Prior-Guided Adversarial Initialization for Fast Adversarial Training2022
- Adversarial Texture for Fooling Person Detectors in the Physical World2022
- Subspace Adversarial Training2022
- A 3D GAN for Improved Large-pose Facial Recognition2021
- Efficient Adversarial Training With Transferable Adversarial Examples2020
- Towards Robust Neural Networks via Random Self-ensemble2018
- Identity Mappings in Deep Residual Networks2016
- Region-Based Convolutional Networks for Accurate Object Detection and SegmentationIEEE Transactions on Pattern Analysis and Machine Intelligence, 2015
- Improvements to Deep Convolutional Neural Networks for LVCSR2013
- Domain Adaptation via Transfer Component AnalysisIEEE Transactions on Neural Networks, 2010