Swin Meets EfficientNet: Lightweight Architectures for GAN-Based Face Forensics

Authors: Sejuti Basu, Ashima Sood, Vijay Kumar, Sahil Sharma

Published: 2026-09-01 18:19:32+00:00

Comment: 12 pages, 2 figures, 1 table. Presented at the International Conference on Computational Techniques in Data Science (IICTDS 2025), online. Proceedings forthcoming

AI Summary

This research investigates efficient detection methods for GAN-generated synthetic faces using image analysis alone, proposing Swin-Transformer-based architectures. The study explores a compact Swin Transformer, ImageNet-1K pre-trained Swin-Tiny and Swin-Small models, and a novel hybrid combining EfficientNet-B0 with a Swin Transformer backend. The EfficientNetB0+Swin hybrid achieved 99% accuracy and 99.44% recall, outperforming other variants and a CNN-only baseline, suggesting that combining hierarchical CNN features with shifted-window self-attention is an efficient method for detecting GAN-generated synthetic faces.

Abstract

Modern generative models, such as GANs, diffusion architectures, and autoregressive systems, now produce facial images that are nearly indistinguishable from authentic photographs. This capability makes detecting forged images increasingly difficult, raising serious concerns about identity theft, fraud, and misinformation campaigns. Our research focuses specifically on GAN-generated synthetic faces, which underpin many face-centric deepfakes, and investigates efficient detection approaches using image analysis alone. Existing detection systems rely heavily on either convolutional neural networks (CNNs) or global vision transformers. While CNNs excel at identifying texture-based local features, they struggle with broader contextual understanding. Traditional Vision Transformer (ViT) models can capture long-range structures effectively, but demand substantial computational resources. Our work explores Swin-Transformer-based architectures across three implementations: a compact Swin Transformer trained from the ground up, ImageNet-1K pre-trained Swin-Tiny and Swin-Small models adapted for binary classification, and a novel hybrid combining EfficientNet-B0's convolutional processing with a Swin Transformer backend. We evaluated all models using the 140K Real and Fake Faces dataset, which includes StyleGAN-generated fake faces alongside authentic images from Flickr and DFDC, with balanced splits for training, validation, and testing. The EfficientNetB0+Swin hybrid achieved 99% accuracy and a 99.44% recall on 5,000 test images, outperforming both pure Swin variants and a previous CNN-only baseline on this dataset. Our results suggest that combining hierarchical CNN features with shifted-window self-attention provides an efficient and computationally lightweight method for detecting GAN-generated synthetic faces.

Key findings

The hybrid EfficientNetB0+Swin architecture achieved the best performance with 99% accuracy and 99.44% recall on the 140K Real and Fake Faces dataset, significantly outperforming pure Swin variants and a CNN-only baseline. This indicates that combining hierarchical CNN features (from EfficientNet) with shifted-window self-attention (from Swin Transformer) provides an efficient and lightweight method for detecting GAN-generated synthetic faces. Swin-Tiny, with ImageNet-1K pretraining, showed better performance than the larger Swin-Small, suggesting that merely increasing model size does not inherently lead to improved performance in this specific task.

Approach

The authors propose three Swin-Transformer-based architectures: a compact Swin Transformer trained from scratch, ImageNet-1K pre-trained Swin-Tiny and Swin-Small models, and a hybrid architecture combining EfficientNet-B0's convolutional processing with a Swin Transformer backend. The hybrid model processes face images with EfficientNet-B0 to extract mid-level and final feature maps, then uses a shallow Swin stack on the mid-level features for contextual understanding, and finally concatenates pooled outputs from both branches for classification.

Datasets

140K Real and Fake Faces dataset (includes StyleGAN-generated faces, authentic images from Flickr and DFDC)

Model(s)

Swin Transformer, Swin-Tiny, Swin-Small, EfficientNet-B0, EfficientNetB0+Swin hybrid

Author countries

India, United Kingdom

← Previous