Research map / Image generation / Autoregressive and latent generation
Discovering Transferable Forensic Features for CNN-Generated Images Detection
Keshigeyan Chandrasegaran, Ngoc-Trung Tran, Alexander Binder, Ngai-Man Cheung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| dropout of top-k forensic features chosen by FF-RS rather than random-k (ablation) | ResNet-50 universal detector · ForenSynths StyleGAN2 | AP: 55.3 vs 98.6 (random-k) | Improved |
| drop top-k T-FF feature maps (ablation) | ResNet-50 universal detector · ForenSynths ProGAN | GAN detection accuracy: 100. -> 3.2 | Hurt |
| drop top-k T-FF feature maps (ablation) | EfficientNet-B0 universal detector · ForenSynths ProGAN | AP: 100. -> 50.0 | Hurt |
| drop low-k feature maps (ablation) | ResNet-50 universal detector · ForenSynths StyleGAN2 | AP: 99.1 -> 99.1 | No clear effect |
| increase dropout size from k=29 to k=228 (ablation) | ResNet-50 universal detector · ForenSynths ProGAN | AP: 98.6 -> 58.6 | Hurt |
| convert counterfeits to grayscale at test time (ablation) | ResNet-50 universal detector · ForenSynths BigGAN | GAN detection accuracy: 85.1 -> 48.8 | Hurt |
| convert counterfeits to grayscale at test time (ablation) | ResNet-50 universal detector · ForenSynths ProGAN | GAN detection accuracy: 100.0 -> 81.5 | Hurt |
| convert counterfeits to grayscale at test time (ablation) | universal detector · ForenSynths unseen GANs | median counterfeit probability: drop by more than 89% | Hurt |
| add random grayscale (50% probability) augmentation during training (CR-universal) | ResNet-50 universal detector · ProGAN, Mood median test | % color-conditional T-FF: 85.1 -> 55.3 | Improved |
| add random grayscale augmentation during training (CR-universal) | EfficientNet-B0 universal detector · ProGAN, Mood median test | % color-conditional T-FF: 51.9 -> 20.0 | Improved |
| add random grayscale augmentation during training (CR-universal) | universal detector · ForenSynths unseen GANs | robustness to color ablation | Improved |
| use pixel-wise attribution (Guided-GradCAM, LRP) to find forensic features | universal detector · ProGAN-trained detector | qualitative discovery of T-FF | No clear effect |
Closest papers by meaning
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.85)
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.84)
- BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors (NeurIPS 2026 · similarity 0.84)
- ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization (NeurIPS 2025 · similarity 0.84)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.84)
- A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
