Research map / Detection and segmentation / Open-set and few-shot detection
DeepfakeImpact: A Two-Stage Benchmark with Real-World Impact in Deepfake Detection
Chaoyu Gong, Han Zhang, Siqiang Luo
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 |
|---|---|---|---|
| start from pre-trained backbone weights (ablation) | SSTGNN (ResNet-50) · SMI-critical dataset | Acc-SMI: 0.705 -> 0.741 | Improved |
| start from pre-trained backbone weights (ablation) | FaceXray (ResNet-50) · SMI-critical dataset | Acc-SMI: 0.247 -> 0.304 | Improved |
| start from pre-trained backbone weights (ablation) | SPSL (Xception) · SMI-critical dataset | Acc-SMI: 0.439 -> 0.462 | Improved |
| swap backbone ResNet-34 to ResNet-50 (ablation) | SSTGNN · SMI-critical dataset | Acc-SMI: 0.729 -> 0.741 | Improved |
| swap backbone among pre-trained ResNet-34/50, EfficientNet-B4, Xception (ablation) | FaceXray, Core, SPSL · SMI-critical dataset | Acc-SMI: FaceXray 0.304 for both ResNets | No clear effect |
| add data augmentations (vs removing all) (ablation) | frame-level detectors (XceptionNet, SIA, RECCE, Effort) · SMI-critical dataset | relative change in Acc-SMI: XceptionNet -2.20% without augmentations | Improved |
| add data augmentations (vs removing all) (ablation) | video-level detectors (SSTGNN, STIL, FTCN, TALL) · SMI-critical dataset | relative change in Acc-SMI: SSTGNN +2.40%, STIL +4.50% when removed | Hurt |
| add Gaussian blur augmentation (ablation) | frame-level detectors · SMI-critical dataset | relative change in Acc-SMI: removal gives +0.40% (SIA), +0.80% (XceptionNet) | Hurt |
| add isotropic scaling augmentation (ablation) | video-level detectors · SMI-critical dataset | relative change in Acc-SMI: removal gives -1.20% (STIL), -0.90% (SSTGNN) | Improved |
| train with SMI-weighted cross-entropy instead of plain cross-entropy (ablation) | SSTGNN · FF++ (t-SNE) | feature separation (qualitative) | Improved |
| use video-level temporal modeling instead of frame-level on face forgery | video-level vs frame-level baselines · Face forgery cross-domain | top-3 AUC: SSTGNN 0.8912 vs UIA-ViT 0.8040 (table) | Improved |
| use video-level temporal modeling on AIG data | video-level vs frame-level baselines · AIG datasets cross-domain | top-3 AUC: SSTGNN 0.8591 vs Effort 0.8583 | No clear effect |
Closest papers by meaning
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.88)
- Through the Lens: Benchmarking Deepfake Detectors Against Moiré-Induced Distortions (NeurIPS 2025 · similarity 0.88)
- ODDN: Addressing Unpaired Data Challenges in Open-World Deepfake Detection on Online Social Networks (AAAI 2025 · similarity 0.88)
- TriDF: Evaluating Perception, Detection, and Hallucination for Interpretable DeepFake Detection (CVPR 2026 · similarity 0.87)
- Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content (CVPR 2025 · similarity 0.87)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.87)
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.
