Research map / Image generation / Autoregressive and latent generation
Learning Self-Consistency for Deepfake Detection
Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, Wei Xia
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 pairwise self-consistency loss (lambda=10 vs lambda=0) to I2G training (ablation) | ResNet-34 with I2G · DFR/CD2/DFDC/DFDC-P avg (cross-dataset) | AUC (%): 73.74 -> 82.83 | Improved |
| add consistency loss (lambda=10 vs 0) on DFDC (ablation) | ResNet-34 with I2G · DFDC (cross-dataset) | AUC (%): 51.72 -> 67.52 | Improved |
| raise lambda from 10 to 100 (ablation) | PCL + I2G · DFR/CD2/DFDC/DFDC-P avg (cross-dataset) | AUC (%): 82.83 -> 82.09 | Mixed |
| add I2G augmentation when training on DF (ablation) | PCL · DFR/CD2/DFDC/DFDC-P avg (trained on DF) | AUC (%): 79.19 -> 86.37 | Improved |
| add I2G augmentation when training on DFDC-P (ablation) | PCL · DFR/CD2/DFDC/DFDC-P avg (trained on DFDC-P) | AUC (%): 74.78 -> 86.35 | Improved |
| add I2G augmentation when training on DFDC-P (ablation) | PCL · DFDC-P (trained on DFDC-P) | AUC (%): 95.53 -> 94.38 | Hurt |
| use I2G only as pretraining (then finetune on DF) instead of joint training (ablation) | PCL · DFR/CD2/DFDC/DFDC-P avg (trained on DF) | AUC (%): 79.19 -> 84.89 | Improved |
| use joint training with I2G instead of I2G pretraining only (ablation) | PCL · DFR/CD2/DFDC/DFDC-P avg (trained on DF) | AUC (%): 84.89 -> 86.37 | Improved |
| patch size 16x16 versus 4x4 (ablation) | PCL + I2G · FF++ (cross-dataset) | AUC (%): 98.32 -> 99.11 | Improved |
| patch size 32x32 versus 16x16 (ablation) | PCL + I2G · FF++ (cross-dataset) | AUC (%): 99.11 -> 98.74 | Hurt |
| deepen backbone from ResNet-18 to ResNet-34 (ablation) | PCL + I2G · DFR/CD2/DFDC/DFDC-P avg (cross-dataset) | AUC (%): 75.74 -> 82.83 | Improved |
| deepen backbone from ResNet-34 to ResNet-50 (ablation) | PCL + I2G · DFR/CD2/DFDC/DFDC-P avg (cross-dataset) | AUC (%): 82.83 -> 83.90 | Improved |
Closest papers by meaning
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.88)
- BPL: Generalizable Deepfake Detection via Bias-only Pair-aware Learning (ICML 2026 · similarity 0.87)
- Hierarchical Contrastive Inconsistency Learning for Deepfake Video Detection (ECCV 2022 · similarity 0.87)
- From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes (NeurIPS 2025 · similarity 0.87)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.86)
- Contrastive Learning for DeepFake Classification and Localization via Multi-Label Ranking (CVPR 2024 · 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.
