Research map / Image generation / GANs
Leveraging Frequency Analysis for Deep Fake Image Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Classify DCT-transformed rather than raw-pixel images | Ridge regression on FFHQ and StyleGAN images · FFHQ vs StyleGAN | accuracy: 75.78\,\% -> 100.00\,\% (+ 24.22\,\%) | Helped |
| Classify DCT coefficients instead of pixels | Source-identification CNN · LSUN bedrooms | accuracy: 98.95\,\% -> 99.64\,\% (+ 0.69\,\%) | Helped |
| Classify DCT coefficients instead of pixels | Source-identification CNN · CelebA | accuracy: 97.80\,\% -> 99.07\,\% (+ 1.27\,\%) | Mixed |
| Train the shallow CNN on DCT coefficients rather than pixels | Source-identification CNN with 20% of training data · LSUN bedrooms | accuracy: 85.82\,\% -> 99.14\,\% | Helped |
| Train the shallow CNN on DCT coefficients rather than pixels | Source-identification CNN with 20% of training data · CelebA | accuracy: 96.33\,\% -> 98.47\,\% | Helped |
| Replace bilinear upsampling with nearest-neighbor upsampling (ablation) | StyleGAN · StyleGAN-generated images, 256 × 256 | DCT ridge-regression accuracy: 85.96\,\% -> 98.24\,\% | Helped |
| Replace bilinear upsampling with binomial upsampling (ablation) | StyleGAN · StyleGAN-generated images, 256 × 256 | DCT ridge-regression accuracy: 85.96\,\% -> 84.20\,\% | No clear effect |
| Use more elaborate upsampling with stronger anti-aliasing (ablation) | StyleGAN · StyleGAN-generated images, 256 × 256 | DCT ridge-regression accuracy: 98.24\,\% -> 85.96\,\% -> 84.20\,\% | Hurt |
| Use DCT rather than pixel input for clean-trained classification under blur | CNN classifier · LSUN bedrooms, blur perturbation (CD) | accuracy: 60.56\,\% -> 61.42\,\% | Helped |
| Use DCT rather than pixel input for clean-trained classification under cropping | CNN classifier · LSUN bedrooms, cropped perturbation (CD) | accuracy: 74.49\,\% -> 83.52\,\% | Helped |
| Use DCT rather than pixel input for clean-trained classification under compression | CNN classifier · LSUN bedrooms, compression perturbation (CD) | accuracy: 68.66\,\% -> 71.86\,\% | Helped |
| Use DCT rather than pixel input for clean-trained classification under noise | CNN classifier · LSUN bedrooms, noise perturbation (CD) | accuracy: 59.51\,\% -> 48.99\,\% | Hurt |
Closest papers by meaning
- Think Twice Before Detecting GAN-Generated Fake Images From Their Spectral Domain Imprints (CVPR 2022 · similarity 0.92)
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.88)
- Transfer Learning of Real Image Features with Soft Contrastive Loss for Fake Image Detection (AAAI 2025 · similarity 0.87)
- Secret Lies in Color: Enhancing AI-Generated Images Detection with Color Distribution Analysis (CVPR 2025 · similarity 0.86)
- FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge (NeurIPS 2024 · similarity 0.86)
- Forgery-aware Adaptive Transformer for Generalizable Synthetic Image Detection (CVPR 2024 · similarity 0.85)
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.