Research map / Video / Video segmentation and tracking
Omni-Fake: Benchmarking Unified Multimodal Social Media Deepfake Detection
Tianxiao Li, Zhenglin Huang, Haiquan Wen, Yiwei He, Xinze Li, Bingyu Zhu, Wuhui Duan, Congang Chen, Zeyu Fu, Yi Dong, Baoyuan Wu, Xiangtai Li and 1 more
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 |
|---|---|---|---|
| use curriculum SFT with modal replay instead of full-mix SFT (ablation) | Omni-Fake-R1 SFT · Omni-Fake-Set val (Image) | detection accuracy: 79.29 -> 87.79 | Improved |
| use curriculum SFT instead of single-modality SFT (ablation) | Omni-Fake-R1 SFT · Omni-Fake-Set val (Image) | detection accuracy: 83.76 -> 87.79 | Improved |
| use curriculum SFT instead of full-mix SFT (ablation) | Omni-Fake-R1 SFT · Omni-Fake-Set val (Audio) | detection accuracy: 71.96 -> 88.45 | Improved |
| use curriculum SFT instead of full-mix SFT (ablation) | Omni-Fake-R1 SFT · Omni-Fake-Set val (Video) | localization IoU: 24.69 -> 38.25 | Improved |
| add unified GSPO RL on top of curriculum SFT (ablation) | Omni-Fake-R1 · Omni-Fake-Set val (Audio) | detection accuracy: 88.45 -> 92.13 | Improved |
| add unified GSPO RL on top of curriculum SFT (ablation) | Omni-Fake-R1 · Omni-Fake-Set val (Image) | localization F1: 43.94 -> 51.63 | Improved |
| add unified GSPO RL on top of curriculum SFT (ablation) | Omni-Fake-R1 · Omni-Fake-Set val (AV-TH) | detection accuracy: 91.33 -> 96.18 | Improved |
| apply GSPO RL without curriculum SFT (RL-only) (ablation) | Omni-Fake-R1 · Omni-Fake-Set val | detection and localization | Hurt |
| increase replay ratio from 0% to 15% (ablation) | curriculum SFT · Audio->Image curriculum | average detection accuracy: 70.19 -> 80.45 | Improved |
| increase replay ratio from 15% to 30% (ablation) | curriculum SFT · Audio->Image curriculum | average IoU: 39.49 -> 37.31 | Hurt |
| remove explanation terms from GSPO reward (ablation) | Omni-Fake-R1 · Omni-Fake-Set | CSS | Improved |
| apply JPEG 70 compression to OOD inputs | Omni-Fake-R1 · Omni-Fake-OOD corruptions (modality average) | detection accuracy: 92.52 -> 89.92 | Hurt |
Closest papers by meaning
- SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model (CVPR 2025 · similarity 0.88)
- Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning (CVPR 2025 · similarity 0.87)
- Lips Are Lying: Spotting the Temporal Inconsistency between Audio and Visual in Lip-Syncing DeepFakes (NeurIPS 2024 · similarity 0.87)
- GLCF: A Global-Local Multimodal Coherence Analysis Framework for Talking Face Generation Detection (AAAI 2025 · similarity 0.86)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.86)
- I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.86)
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.
