Wavelet-Guided Frequency-Domain Adaptive Learning: Balancing Adversarial Defense and High-Fidelity Image Reconstruction
- 30 October 2026
- journal article
- research article
- Published by World Scientific Pub Co Pte Ltd in International Journal of Pattern Recognition and Artificial Intelligence
Abstract
Deep learning has achieved remarkable success across diverse domains, including image recognition, segmentation, and restoration. However, adversarial attacks — where imperceptible perturbations are introduced to normal images to mislead models — pose significant security risks to deep neural networks, underscoring the critical need for robust adversarial defense strategies. Image reconstruction-based defense is widely adopted for its downstream model independence and flexibility, yet existing methods often prioritize defense effectiveness at the expense of image quality, limiting their applicability in high-stakes scenarios such as medical diagnosis and remote sensing. To address this imbalance, we propose WS-Net, a wavelet-guided frequency-domain adaptive denoising network. Leveraging the nonuniform distribution of adversarial noise, WS-Net decomposes images into low-frequency and high-frequency components via Discrete Wavelet Transform (DWT), applying denoising exclusively to high-frequency components to preserve critical structural information. At its core is the Swin Network Denoiser (SND), constructed with Swin Transformer-based blocks that utilize shifted window self-attention to accurately distinguish noise from image details. A multi-level loss function — combining pixel-level [Formula: see text] loss, feature-level perceptual loss, style loss, and adversarial loss — constrains the model from multiple dimensions. Experimental results on the ImageNet dataset demonstrate that WS-Net outperforms state-of-the-art models, achieving a top-1 accuracy of 71.4% on clean images, 70.2% under Gaussian noise, and superior defense performance across nine adversarial attacks (e.g. 68.5% accuracy against CW attacks, 67.9% against DeepFool). Notably, WS-Net achieves a Peak Signal-to-Noise Ratio (PSNR) of 30.2[Formula: see text]dB and a Structural Similarity Index (SSIM) of 0.87, balancing defense robustness and image fidelity effectively. The code for WS-Net is available at https://github.com/faaatwang/WS_Net.git .Keywords
Funding Information
- Key Technologies Research and Development Program (2020YFB1712401)
- Henan Provincial Science and Technology Research Project (242102221017)
- Henan Provincial Science and Technology Research Project (232102210090)
- Key Scientific Research Project of Colleges and Universities in Henan Province (23A520015)
- Key Special Technologies Research and Development Program in Henan Province (231111211900)
- Graduate Student Joint Training Base Project in Henan Province (YJS2023JD04)
This publication has 38 references indexed in Scilit:
- GCSTormer: Gated swin transformer with channel weights for image denoisingExpert Systems with Applications, 2025
- Adversarial purification of information maskingNeurocomputing, 2024
- Practical Blind Image Denoising via Swin-Conv-UNet and Data SynthesisMachine Intelligence Research, 2023
- Defense Against Adversarial Attacks with Efficient Frequency-Adaptive Compression and ReconstructionPattern Recognition, 2023
- Feature-filter: Detecting adversarial examples by filtering out recessive featuresApplied Soft Computing, 2022
- PyTorch2021
- Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search2020
- Deep Learning for Generic Object Detection: A SurveyInternational Journal of Computer Vision, 2019
- Adversarial image detection in deep neural networksMultimedia Tools and Applications, 2018
- Perceptual Losses for Real-Time Style Transfer and Super-Resolution2016