Research map / Video / Video segmentation and tracking
Two-branch Recurrent Network for Isolating Deepfakes in Videos
Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, Wael AbdAlmageed
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 |
|---|---|---|---|
| add per-layer learning rates, dropout and the new loss to the encoder (ablation) | DenseNet encoder · FF++ Deepfakes c40 (no LSTM) | tAUC1%: 57.63 -> 76.04 | Improved |
| add per-layer learning rates, dropout and the new loss to the encoder (ablation) | DenseNet encoder · FF++ Deepfakes c40 (no LSTM) | TAR1%: 81.50 -> 92.84 | Improved |
| add per-layer learning rates, dropout and the new loss to the encoder (ablation) | DenseNet encoder · FF++ Deepfakes c40 (no LSTM) | tAUC10%: 86.61 -> 89.91 | Improved |
| add the second (LoG) branch to the single RGB branch (ablation) | backbone without LSTM · FF++ Deepfakes c23 | tAUC1%: 56.78 -> 61.70 | Improved |
| add the second (LoG) branch to the single RGB branch (ablation) | backbone without LSTM · FF++ Deepfakes c23 | tAUC10%: 61.14 -> 70.80 | Improved |
| add the second (LoG) branch to the single RGB branch (ablation) | backbone without LSTM · FF++ Deepfakes c23 | TAR1%: 99.34 -> 98.34 | Hurt |
| add dropout/fine-tuning schedule to the LSTM model (ablation) | LSTM model · FF++ Deepfakes c40 (with LSTM) | tAUC1%: 57.21 -> 73.58 | Improved |
| concatenate (cat) bi-directional LSTM outputs instead of summing (ablation) | LSTM model · FF++ Deepfakes c40 (with LSTM) | tAUC1%: 67.49 -> 73.58 | Improved |
| use 128 instead of 256 LSTM hidden nodes (g=2) (ablation) | LSTM model · FF++ Deepfakes c40 (with LSTM) | tAUC1%: 76.35 -> 81.53 | Improved |
| split fusion conv1x1 filters into two groups (g=2 instead of g=1) (ablation) | LSTM model · FF++ Deepfakes c40 (with LSTM) | tAUC1%: 73.58 -> 81.53 | Improved |
| use proposed model instead of Xception at medium compression | FF++ c23 · FaceForensics++ c23 video-level | AUC: 92.50 -> 99.12 | Improved |
| use proposed model instead of Xception at c40 for frame-level tAUC10% and pAUC10% | FF++ c40 · FaceForensics++ c40 frame-level | tAUC10%: 45.92 -> 40.41 | Hurt |
Closest papers by meaning
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.87)
- FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge (NeurIPS 2024 · similarity 0.87)
- Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content (CVPR 2025 · similarity 0.87)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.87)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.86)
- DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion (NeurIPS 2024 · 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.
