Quang Huy Nguyen and Van Huy Pham
- [03/2026] Our presentation achieved the Outstanding Presentation Award.
- [01/2026] This research was accepted and selected for an oral presentation at The Digital Convergence in Economics, Society and Technology (DCEST 2026).
The rapid advancement of Generative Adversarial Networks (GANs) continues to challenge digital media integrity, while purely spatial-domain forensic models remain sensitive to overfitting and weakly generalizable artifacts. To address this, we propose the Adaptive Spectral-Band Weighting (ASBW) layer, a lightweight frequency-domain module that applies the Discrete Fourier Transform (DFT) and a compact Multi-Layer Perceptron (MLP) to reweight discriminative spectral cues caused by GAN generation. We evaluate ASBW insertion at different positions across the ResNet-18 pipeline (after each convolutional block), and compare these placement strategies with the baseline CNN. Results on the real/fake face benchmark show that ASBW effectiveness is strongly position-dependent: the full configuration achieves the best performance (93.12%), while placing ASBW after later blocks (e.g., Position 3: 92.22%) consistently improves over the baseline (89.77%). These findings highlight that frequency-aware amplification is most beneficial when integrated with suitable feature hierarchy depth, and provide practical guidance for architecture design in GAN image forensics.
If you find this work useful, please cite:
@inproceedings{nguyen2026asbw,
title = {ASBW: A Frequency-Domain Analysis Approach for Distinguishing GAN-Generated Images from Real Images},
author = {Nguyen, Quang Huy and Pham, Van Huy},
booktitle = {In Proceeding of the Digital Convergence in Economics, Society and Technology (DCEST) 2026 International Conference},
year = {2026},
pages = {95-105}
}- Quang Huy Nguyen: nguyenquanghuy.st@tdtu.edu.vn
- Van Huy Pham: huy.phamvan@umt.edu.vn