Research map / Continual learning / Online and replay-based
Stacking Brick by Brick: Aligned Feature Isolation for Incremental Face Forgery 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 SUR-LID, an incremental face forgery detector that uses uniform-sparse replay plus isolation loss, distribution re-filling and decision alignment to reduce catastrophic forgetting, evaluated on a new two-protocol IFFD benchmark.
Problem
Catastrophic forgetting in incremental face forgery detection, where treating all forgeries as one "Fake" class causes distributions of different forgery types to override one another.
Approach
Replay a uniformly sparse, stability-filtered subset that approximates each previous global feature distribution (Sparse Uniform Replay), then isolate real/fake and task-wise latent distributions with a supervised contrastive loss over replayed and re-filled features, and align linear decision boundaries across tasks via a normalized weight update.
Key results
- SUR-LID reaches 0.9315 average AUC on Protocol 1 at task 4, versus 0.8752 for HDP and 0.8606 for CoReD.
- On Protocol 2 at task 4, SUR-LID reaches 0.9433 average AUC versus 0.7826 for HDP.
- Ablation: removing the isolation loss drops average AUC from 0.9315 to 0.8841, and removing IDA drops it to 0.8352.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Incremental Decision Alignment (IDA) (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.8352 -> 0.9315 | Helped |
| add isolation loss (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.8841 -> 0.9315 | Helped |
| add Distribution Re-filling (DR) (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.9160 -> 0.9315 | Helped |
| add all proposed components (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.8217 -> 0.9315 | Helped |
| replace Center replay with SUR (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.8591 -> 0.9315 | Helped |
| replace Center+Hard replay with SUR (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.8776 -> 0.9315 | Helped |
| replace Random replay with SUR (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.7826 -> 0.9315 | Helped |
| replace Random Uniform replay with SUR (ablation) | SUR-LID · Protocol 1 T4 average | AUC: 0.9069 -> 0.9315 | Helped |
| use SUR-LID instead of DFIL | Incremental face forgery detection · Protocol 1 T4 average | AUC: 0.8628 -> 0.9315 | Helped |
| use SUR-LID instead of CoReD | Incremental face forgery detection · Protocol 2 T4 average | AUC: 0.8141 -> 0.9433 | Helped |
Surprising findings
- Existing IFFD methods perform poorly on Protocol 2 (forgery-type incremental with shared real domain), where forgery-irrelevant features are consistent across forgeries and get overridden.
Limitations
- Replay buffer size and stability/uniformity approximations are heuristic; the exact optimal selection is replaced by a local approximation.
- Evaluated only on face forgery datasets with EfficientNet-B4 backbone.
Experimental setup
Datasets: FaceForensics++ Celeb-DF-v2 DFDC Preview DeepFakeDetection DF40 (MCNet, BlendFace, StyleGAN3) DiffusionFace (SDv21)
Benchmarks: Protocol 1 (dataset incremental) Protocol 2 (forgery type incremental) Protocol 3 (classical FF++/DFDCP/DFD/CDF, supplementary)
Metrics: AUC ACC
Baselines: LwF iCaRL DER CoReD DFIL HDP Lower Bound
Builds on: DFIL DeepFakeBench iCaRL supervised contrastive loss
Closest papers by meaning
- DevFD : Developmental Face Forgery Detection by Learning Shared and Orthogonal LoRA Subspaces (NeurIPS 2025 · similarity 0.88)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.88)
- Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector (CVPR 2025 · similarity 0.88)
- Critical Forgetting-Based Multi-Scale Disentanglement for Deepfake Detection (AAAI 2025 · similarity 0.87)
- FakeDiffer: Distributional Disparity Learning on Differentiated Reconstruction for Face Forgery Detection (AAAI 2025 · similarity 0.87)
- A Hitchhiker's Guide to Fine-Grained Face Forgery Detection Using Common Sense Reasoning (NeurIPS 2024 · similarity 0.87)
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.