Research map / Detection and segmentation / CNN features
Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection
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.
Proposes an information-bottleneck-based deepfake detector that extracts multiple non-overlapping local forgery features (ResNet-34 blocks) and fuses them via Local and Global Information Losses, reporting state-of-the-art in-dataset and cross-dataset AUC on FF++, Celeb-DF and DFDC.
Problem
Deepfake detectors overfit to one or a few local forgery regions and generalize poorly to unseen datasets, and neural feature extractors lack theoretical guarantees that they capture sufficient task-relevant information while discarding superfluous information.
Approach
Extract several local features with separate ResNet-34 blocks, enforce orthogonality and label comprehensiveness with a variational Local Information Loss derived from interaction information, then fuse them into a global feature with a Global Information Loss motivated by the information bottleneck.
Key results
- In-dataset frame-level AUC of 0.983 on FF++ (C23), 0.999 on Celeb-DF-V2, and 0.939 on DFDC.
- Trained on FF++, achieves frame-level AUC 0.818 on CD1, 0.864 on CD2, 0.851 on DFDC-P, and 0.721 on DFDC, versus prior best ICT at 0.814 and 0.857 on CD1/CD2.
- Ablation on CD2: removing both losses drops AUC from 0.864 to 0.816.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Local Information Loss and Global Information Loss (ablation) | Ours (ResNet-34) · FF++ (C23) | AUC: 0.966 to 0.983 | Helped |
| add Local Information Loss and Global Information Loss (ablation) | Ours (ResNet-34) · Celeb-DF-V2 | AUC: 0.816 to 0.864 | Helped |
| add Local Information Loss while retaining Global Information Loss (ablation) | Ours (ResNet-34) · FF++ (C23) | AUC: 0.979 to 0.983 | Helped |
| add Local Information Loss while retaining Global Information Loss (ablation) | Ours (ResNet-34) · Celeb-DF-V2 | AUC: 0.840 to 0.864 | Helped |
| add Global Information Loss while retaining Local Information Loss (ablation) | Ours (ResNet-34) · FF++ (C23) | AUC: 0.977 to 0.983 | Helped |
| add Global Information Loss while retaining Local Information Loss (ablation) | Ours (ResNet-34) · Celeb-DF-V2 | AUC: 0.827 to 0.864 | Helped |
| replace ResNet-34 with MobileNet-V1 in local information blocks (ablation) | Ours · Celeb-DF-V2 | AUC: 0.864 vs 0.858 | No clear effect |
| replace ResNet-34 with EfficientNet-B0 in local information blocks (ablation) | Ours · Celeb-DF-V2 | AUC: 0.864 vs 0.863 | No clear effect |
| replace adaptive local feature extraction with evenly cropped regions (ablation) | Ours · FF++ (C23) | AUC: 0.983 vs 0.959 | Hurt |
| replace adaptive local feature extraction with evenly cropped regions (ablation) | Ours · Celeb-DF-V2 | AUC: 0.864 vs 0.798 | Hurt |
| train on FF++ and test on in-the-wild videos | Ours versus Xception · In-the-wild videos | AUC: 0.579 to 0.728 | Helped |
Surprising findings
- Simply cropping the face into evenly spaced non-overlapping regions performs worse (CD2 AUC 0.798 vs 0.864 for the proposed method), suggesting learned disentanglement matters more than fixed partitioning.
- Performance peaks at four local blocks and declines at seven, indicating overfitting from excess parameters.
Limitations
- Purely data-driven; the authors note that prior knowledge could be more effective for some classes or forgeries.
- Performance is sensitive to the number of local blocks, with generalization declining at n=7.
- Theoretical lower bound and global loss rely on approximations, and the Global Information Loss is derived only partially, with the superfluous-information term left to network operations.
Experimental setup
Datasets: FaceForensics++ (C23) Celeb-DF-V1 Celeb-DF-V2 DFDC-Preview DFDC
Benchmarks: FaceForensics++ (C23) Celeb-DF-V1 Celeb-DF-V2 DFDC-Preview DFDC
Metrics: frame-level AUC video-level AUC ACC LogLoss
Baselines: Xception Face X-ray SBIs ICT F3-Net AUNet OST RECCE Multi-Attentional LipForensics
Builds on: information bottleneck interaction information Multi-Attentional deepfake detection ResNet
Closest papers by meaning
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.91)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.90)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.90)
- Critical Forgetting-Based Multi-Scale Disentanglement for Deepfake Detection (AAAI 2025 · similarity 0.89)
- FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge (NeurIPS 2024 · similarity 0.89)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.89)
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.