Research map / Label-efficient and robust learning / OOD and anomaly detection
Transcending Forgery Specificity with Latent Space Augmentation for Generalizable 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 LSDA, a distillation-based deepfake detector that augments forgery features in latent space (within-domain centrifugal/affine/additive and cross-domain Mixup) and uses a pre-trained ArcFace real encoder to improve cross-dataset generalization.
Problem
Deepfake detectors overfit to forgery-specific artifacts and generalize poorly to unseen forgery types and datasets when train and test distributions differ.
Approach
Enlarge the forgery feature space in latent space rather than pixel space: within-domain transformations push features toward hard examples and perturb them, cross-domain Mixup interpolates features of different forgery types, and a student encoder is distilled from per-domain teacher encoders (with a frozen ArcFace real teacher).
Key results
- Average frame-level AUC of 0.826 across CDF-v1, CDF-v2, DFD, DFDC, DFDCP on DeepfakeBench, a ~5.9% gain over the best listed baseline (UCF at 0.763).
- Video-level AUC of 0.911 on CDF-v2 and 0.770 on DFDC, outperforming listed SOTA methods (e.g., SBI* 0.906 and 0.724).
- On GAN/diffusion-generated image detection, LSDA reaches 0.810 (StarGAN), 0.854 (DDPM), 0.748 (DDIM) versus SBI's 0.787, 0.744, 0.648.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add within-domain and cross-domain latent augmentation (ablation) | EfficientNet-B4 deepfake detector · Average across CDF-v1, CDF-v2, DFDCP, DFDC | AUC: 0.755 -> 0.825 | Helped |
| add cross-domain latent augmentation (ablation) | EfficientNet-B4 deepfake detector · Average across CDF-v1, CDF-v2, DFDCP, DFDC | AUC: 0.755 -> 0.821 | Helped |
| add within-domain latent augmentation (ablation) | EfficientNet-B4 deepfake detector · Average across CDF-v1, CDF-v2, DFDCP, DFDC | AUC: 0.755 -> 0.819 | Helped |
| use ArcFace-pretrained real encoder instead of EFNB4 real encoder (ablation) | LSDA · Average across CDF-v1, CDF-v2, DFDCP, DFDC | AUC: 0.804 -> 0.812 | Helped |
| use ArcFace-pretrained real encoder instead of iResNet101 without face-recognition pretraining (ablation) | LSDA · Average across CDF-v1, CDF-v2, DFDCP, DFDC | AUC: 0.790 -> 0.812 | Helped |
| apply LSDA rather than the strongest listed baseline | deepfake detector trained on FF++ c23 · CDF-v1 | frame-level AUC: 0.815 -> 0.867 | Helped |
| apply LSDA rather than the strongest listed baseline | deepfake detector trained on FF++ c23 · CDF-v2 | frame-level AUC: 0.765 -> 0.830 | Helped |
| apply LSDA rather than the strongest listed baseline | deepfake detector trained on FF++ c23 · DFD | frame-level AUC: 0.816 -> 0.880 | Helped |
| apply LSDA rather than the strongest listed baseline | deepfake detector trained on FF++ c23 · DFDC | frame-level AUC: 0.719 -> 0.736 | Helped |
| apply LSDA rather than the strongest listed baseline | deepfake detector trained on FF++ c23 · DFDCP | frame-level AUC: 0.759 -> 0.815 | Helped |
| apply LSDA instead of SBI | deepfake detector trained on FF++ c23 · Stable Diffusion-generated images | AUC: 0.478 -> 0.506 | Helped |
Surprising findings
- Using ArcFace as the real-class encoder markedly improves results over EfficientNet-B4 or iResNet101 for the real encoder, indicating face recognition priors help deepfake detection.
Limitations
- Robustness and extensibility advantages are shown only on selected perturbations and a few generators; SD results remain low (0.506).
- Relies on a pre-trained ArcFace real encoder and multiple per-domain teacher encoders, adding design complexity.
- Gains on DFDC are modest relative to some methods (e.g., 0.736 frame-level AUC).
Experimental setup
Datasets: FaceForensics++ c23 DeepfakeDetection (DFD) DFDC DFDCP CelebDF-v1 CelebDF-v2 StarGAN DDPM DDIM Stable Diffusion
Benchmarks: FF++ c23 cross-dataset (CDF-v1, CDF-v2, DFD, DFDC, DFDCP) DeepfakeBench GAN and diffusion generated image detection (StarGAN, DDPM, DDIM, SD)
Metrics: frame-level AUC video-level AUC AP EER
Baselines: Meso4 MesoIncep CNN-Aug Xception EfficientB4 CapsuleNet FWA Face X-ray FFD CORE Recce UCF
Builds on: Mixup SBI ArcFace DeepfakeBench EfficientNet
Closest papers by meaning
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.90)
- FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · similarity 0.90)
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.89)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.89)
- From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes (NeurIPS 2025 · similarity 0.88)
- Can We Leave Deepfake Data Behind in Training Deepfake Detector? (NeurIPS 2024 · similarity 0.88)
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.