Research map / Detection and segmentation / CNN features
LAA-Net: Localized Artifact Attention Network for Quality-Agnostic and 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 LAA-Net, a deepfake detector trained only on real images with pseudo-fakes, combining explicit heatmap and self-consistency attention on vulnerable blending points with an Enhanced FPN, achieving SOTA cross-dataset generalization.
Problem
Detecting high-quality deepfakes with subtle localized blending artifacts, and generalizing to unseen manipulations and datasets, where prior binary classifiers with implicit attention generalize poorly.
Approach
An explicit attention mechanism in a multi-task framework: heatmap regression and self-consistency regression centered on automatically derived "vulnerable points" from blending-boundary masks of pseudo-fakes (no manual annotation, real data only), plus E-FPN that propagates low-level features with redundancy-reducing masking.
Key results
- Ours (w/ SBI) achieves 99.96% in-dataset AUC on FF++ and 80.03% AUC / 81.08% AP on Celeb-DF v2 under cross-dataset protocol (trained on FF++).
- Cross-dataset ablation: full model (heatmap, self-consistency, E-FPN) averages 90.20% AUC over CDF2, DFD, DFDC, DFW versus 74.35% for the baseline without components.
- Outperforms Multi-attentional by 27.14% AUC on CDF2 according to the paper.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SBI instead of BI pseudo-fake synthesis | LAA-Net · CDF2 | AUC (%): 86.28 -> 95.40 | Helped |
| Use SBI instead of BI pseudo-fake synthesis | LAA-Net · CDF2 | AP (%): 56.89 -> 81.08 | Helped |
| Add the heatmap branch (ablation) | LAA-Net with consistency branch and E-FPN · CDF2 | AUC (%): 79.87 -> 95.40 | Helped |
| Add the consistency branch (ablation) | LAA-Net with heatmap branch and E-FPN · CDF2 | AUC (%): 95.56 -> 95.40 | No clear effect |
| Add E-FPN (ablation) | LAA-Net with consistency and heatmap branches · CDF2 | AUC (%): 91.56 -> 95.40 | Helped |
| Replace traditional FPN with E-FPN and integrate F(6), F(5), F(4), F(3), and F(2) (ablation) | LAA-Net · CDF2 | AUC (%): 88.35 -> 95.40 | Helped |
| Set λ1=10 and λ2=100 instead of λ1=1 and λ2=1 (ablation) | LAA-Net · CDF2 | AUC (%): 90.69 -> 95.40 | Helped |
Surprising findings
- Extremely sensitive to Gaussian noise perturbations (53.9% accuracy under noise) despite robustness to color changes.
- Using BI instead of SBI gives slightly better DFD performance (86.15 vs 86.94 reversed; BI 69.69 AUC on DFD is lower), while LAA-Net (w/BI) is marginally superior on DFD only in its own table comparison.
Limitations
- Highly sensitive to structural perturbations such as Gaussian noise.
- Frame-level spatial method; temporal dimension not modeled.
Experimental setup
Datasets: FaceForensics++ Celeb-DF v2 DFD DFDC DFW
Benchmarks: FaceForensics++ Celeb-DF v2 DFD DFDC DFW
Metrics: AUC AP
Baselines: Xception FaceXRay+BI Multi-attentional RECCE SFDG CADDM SBI AltFreezing UCF AUNet PCL+I2G
Builds on: Self-Blended Images Face X-ray Self-consistency learning (PCL+I2G) Feature Pyramid Network EfficientNet
Closest papers by meaning
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.85)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.84)
- From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes (NeurIPS 2025 · similarity 0.84)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.84)
- Generalizing Deepfake Video Detection with Plug-and-Play: Video-Level Blending and Spatiotemporal Adapter Tuning (CVPR 2025 · similarity 0.84)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.84)
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.