Swin Meets EfficientNet: Lightweight Architectures for GAN-Based Face Forensics
Sejuti Basu, Ashima Sood, Vijay Kumar, Sahil Sharma
cs.CV · cs.AI
Quick verdict
Reports 99% acc for an EfficientNet+Swin hybrid on StyleGAN faces; results plausible but the key EfficientNet-only ablation and FLOPs data are missing.
Summary
The paper investigates Swin-Transformer-based detectors for GAN-generated synthetic faces, evaluated on the 140K Real and Fake Faces dataset (StyleGAN fakes vs. Flickr/DFDC real images) using a balanced 25K/5K/5K train/validation/test subset. Three configurations are compared: a compact Swin trained from scratch (88.50% accuracy), ImageNet-pretrained Swin-Tiny (91.13%) and Swin-Small (89.14%), and a hybrid combining an EfficientNet-B0 backbone with a Swin head operating on mid-level feature maps at 384x384 input. The hybrid reaches 99.00% accuracy, 98.57% precision, and 99.44% recall on the 5,000-image test set, roughly 11 points above an ELA+CNN baseline from prior work; Grad-CAM visualizations are used to argue the Swin head contributes contextual reasoning. The authors explicitly confine their claims to GAN-generated still faces, acknowledge the single-dataset/single-generator scope, and call for cross-dataset, cross-generator, and video extensions. Key limitations: the hybrid is the only model run at 384x384 and no EfficientNet-B0-only metrics at that resolution are reported, so the gain cannot be attributed to the Swin head versus resolution or tuning; the repeated 'lightweight' claim is never supported by parameter/FLOP/latency measurements; results are single runs without variance or significance testing; and the 140K benchmark is a 2019-era StyleGAN testbed approaching saturation, with no identity-level split discussion or code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A hybrid EfficientNet-B0+Swin architecture achieves 99.00% accuracy and 99.44% recall on a balanced 5,000-image test split of the 140K Real and Fake Faces dataset, outperforming pure Swin variants and an ELA+CNN baseline by over 11 points.
- ImageNet-1K pretraining improves Swin detection (88.50% to 91.13% accuracy), while increasing backbone size from Swin-Tiny to Swin-Small does not (89.14%).
- A compact Swin Transformer trained from scratch rivals a purpose-designed ELA+CNN baseline (88.50% vs 87.46%) without any pretraining.
- Combining hierarchical CNN features with shifted-window self-attention is a computationally lightweight approach to GAN-face detection.
Strengths
Honest scope discipline
Sections 4.4 and 5 explicitly restrict conclusions to GAN stills on one dataset, note that detectors degrade on unseen manipulations (citing [21]), and frame the 99% result as 'controlled evidence rather than... real-world robustness' — rarer candor than typical for this subfield.
Systematic Swin design-space coverage
Table 1 isolates pretraining (Swin Base 88.50% -> Swin-Tiny 91.13% accuracy) and capacity (Swin-Small 89.14% < Swin-Tiny despite more layers/epochs), supporting the compact-backbone argument with an actual ablation pair.
Fully disclosed configurations
Internally consistent headline numbers
Weaknesses
Missing decisive ablation for the central claim
'Lightweight' claimed but never measured
No statistical treatment
Weak baseline and saturated benchmark
Anomalous training hyperparameter
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper evaluates Swin-Transformer-based architectures for binary detection of GAN-generated faces on the 140K Real and Fake Faces dataset (StyleGAN fakes vs. Flickr/DFDC reals), using a balanced 25K/5K/5K train/validation/test subset. Three configurations are studied (Section 3.3): a compact Swin trained from scratch, ImageNet-1K-pretrained Swin-Tiny and Swin-Small, and a hybrid fusing EfficientNet-B0 with a Swin head applied to mid-level feature maps at 384x384 input. The headline claims: the hybrid reaches 99.00% accuracy and 99.44% recall on 5,000 held-out images (Table 1), beating pure Swin variants and a prior ELA+CNN baseline by over 11 points; pretraining helps Swin but capacity does not; and the design is computationally lightweight. Credit where due: the authors scope their claims carefully — Section 4.4 and Section 5 explicitly call this 'controlled evidence' for one dataset and one generator family, not real-world robustness.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does the EfficientNet+Swin hybrid achieve on the 140K Real and Fake Faces dataset?
The paper reports 99.00% accuracy, 98.57% precision, and 99.44% recall on a balanced test set of 5,000 images (2,500 real / 2,500 StyleGAN fake). This is about 11 percentage points above the ELA+CNN baseline (87.46% accuracy) it compares against, and well above its own Swin-Tiny (91.13%) and Swin-Small (89.14%) variants.
What dataset was used and how was it split?
The 140K Real and Fake Faces dataset: 70,000 real faces from Flickr and the DFDC dataset, and 70,000 StyleGAN-generated fakes. The authors used a balanced subset of 25,000 training, 5,000 validation, and 5,000 test images, each split 50% real / 50% fake, with faces detected and cropped using dlib.
Is code available for this paper?
No code release is mentioned. The paper links the public Kaggle version of the dataset and the HuggingFace checkpoints for Swin-Tiny/Swin-Small, and lists hyperparameters in its Section 4.2, but provides no repository or training scripts for the hybrid model.
Does this detector work on diffusion-model faces or real-world deepfakes?
The paper does not test this. All evaluation uses StyleGAN-generated faces from a single dataset; the authors explicitly state the results are controlled-condition evidence only and list cross-generator, cross-dataset, and video evaluation as future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1 and 2 are image content absent from the extracted text; assessments of them rely solely on the paper's in-text descriptions (e.g., the 99.8% peak validation accuracy and Grad-CAM commentary in Section 4.2/4.3).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.