Research map / Detection and segmentation / Open-set and few-shot detection
Through the Lens: Benchmarking Deepfake Detectors Against Moiré-Induced Distortions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds DeepMoiréFake (DMF), a 12,832-video dataset of deepfake videos re-captured from screens with smartphones, and benchmarks 15 deepfake detectors plus demoiréing, denoising, and deblurring defenses under authentic and synthetic Moiré and compression.
Problem
Deepfake detectors are evaluated on clean data, but smartphone recapture of deepfakes from screens introduces Moiré artifacts that may degrade detection; the effect and mitigation via demoiréing are largely unstudied.
Approach
Create a real-world Moiré-impacted deepfake dataset by recapturing videos from five public deepfake datasets on four monitors with two phones under two lighting conditions, then systematically benchmark detectors and restoration-based defenses against authentic Moiré, synthetic Moiré (SMPA-MA, SMPA-SPS), and compression.
Key results
- Moiré artifacts degrade detector AUC by up to 25.4% (average about 10.7%), with the largest drops on LED monitors such as BenQ and LG.
- Demoiréing worsened detection: average AUC drop of 14.7 points relative to original data, up to 17.2 points, while CCViT retained the best average of 79.2%.
- Synthetic Moiré (SMPA-MA) reduced MAT accuracy by 21.4 points, while Rossler and CADDM sometimes improved.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Introduce authentic screen-captured Moiré patterns | Deepfake detectors on DMF · DMF | accuracy: average performance decline of 10.7%; up to 25.4% | Hurt |
| Add synthetic Moiré patterns | MAT · DMF, SMPA | accuracy: 21.4 percentage points | Hurt |
| Add SMPA-MA | Rossler and CADDM · DMF, SMPA-MA | AUC: 78.1 -> 83.1; 73.0 -> 86.8 | Helped |
| Apply demoiréing | Deepfake detectors on DMF · DMF | accuracy: average decline of 6.1%; up to 17.2% | Hurt |
| Apply ESDNet (FHDMi) rather than leaving Moiré in the image | Image-based detectors · DMF | AUC: -1.5 | Mixed |
| Apply ESDNet (UHDM) rather than leaving Moiré in the image | Image-based detectors · DMF | AUC: -7.8 | Hurt |
| Apply video demoiréing | LipForensics · DMF videos | AUC: 87.5 -> 90.6 | Helped |
| Apply video demoiréing | AltFreezing · DMF videos | AUC: 84.4 -> 74.7; 84.4 -> 92.9 | Mixed |
| Apply video demoiréing | FTCN · DMF videos | AUC: 43.8 -> 68.8; 43.8 -> 40.6 | Mixed |
| Change lighting from off to on | LipForensics, AltFreezing, and CCViT · DMF | AUC: 78.7 -> 81.3; 80.6 -> 82.1; 83.5 -> 84.5 | No clear effect |
| Apply Gaussian blur with a 15×15 kernel | MAT · DMF | AUC: 87.0 -> 57.7 | Hurt |
| Apply sharpening kernel value 3 | MAT · DMF | AUC: 87.0 -> 98.2 | Helped |
Surprising findings
- Demoiréing methods, intended as mitigation, reduced detection performance more than the Moiré artifacts themselves (14.7 vs 10.1 point average drop).
- Synthetic Moiré attacks improved Rossler and CADDM performance in some settings instead of degrading it.
Limitations
- Controlled capture setup with limited hardware (four monitors, two phones) and does not cover OLED or 8K displays, motion, or platform filters.
- Dataset release restricted to DOI-based request with non-commercial terms.
Experimental setup
Datasets: FaceForensics++ Celeb-DF DFD DFDC UADFV
Benchmarks: DeepMoiréFake (DMF) FaceForensics++ Celeb-DF DFD DFDC UADFV
Metrics: AUC accuracy precision recall F1-score
Baselines: Rossler MAT CADDM CCViT LipForensics AltFreezing FTCN LRNet
Builds on: FaceForensics++ Celeb-DF DFDC DFD UADFV
Closest papers by meaning
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.88)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.87)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.87)
- From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes (NeurIPS 2025 · similarity 0.86)
- AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (CVPR 2024 · similarity 0.85)
- Towards More General Video-based Deepfake Detection through Facial Component Guided Adaptation for Foundation Model (CVPR 2025 · 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.