Research map / Continual learning / Online and replay-based
Region-Based Optimization in Continual Learning for Audio 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 RegO, a continual learning method for audio deepfake detection that uses Fisher-information-based partitioning of neurons into four regions with region-adaptive gradient projection plus an Ebbinghaus-inspired neuron forgetting mechanism.
Problem
Audio deepfake detectors degrade on diverse, evolving real-world forgeries; continual learning must balance memory stability against learning plasticity when new synthesis methods arrive sequentially.
Approach
Partitions neurons via Fisher information computed separately for real and fake audio into four regions (unimportant, real-only, fake-only, both), applying fine-tuning, parallel projection, orthogonal projection, or sample-proportion-weighted gradients respectively, and periodically releases low-memory neurons using an Ebbinghaus forgetting curve.
Key results
- RegO achieves average EER 6.34% on EVDA versus 8.06% for RWM (the strongest continual baseline), with best results on 7 of 8 experiences.
- On CLEAR image classification, RegO averages higher accuracy than all continual baselines, with the top score on all 10 experiences except tied-second situations, reaching 94.75% on Exp10.
- Ablations show removing region-adaptive optimization raises average EER from 6.34% to 7.63%, and removing the Ebbinghaus mechanism raises it to 6.80%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RegO instead of RWM | Continual audio deepfake detection · EVDA Avg | EER (%): 8.06 -> 6.34; 21.3% improvement | Helped |
| Use RegO instead of EWC | Continual audio deepfake detection · EVDA Avg | EER (%): 7.68 -> 6.34 | Helped |
| Use RegO instead of Replay-All | Continual audio deepfake detection · EVDA Avg | EER (%): 3.61 -> 6.34 | Hurt |
| Use RegO instead of EWC | Continual image recognition · CLEAR Exp_10 | Accuracy (%): 93.13 -> 94.75 | Helped |
| Use RegO instead of Replay-All | Continual image recognition · CLEAR Exp_10 | Accuracy (%): 95.56 -> 94.75 | Hurt |
| Add Ebbinghaus Forgetting Mechanism (EFM) (ablation) | RegO · EVDA Avg | EER (%): 6.80 -> 6.34 | Helped |
| Add Importance Region Localization (IRL) (ablation) | RegO · EVDA Avg | EER (%): 7.38 -> 6.34 | Helped |
| Add Region-Adaptive Optimization (RAO) instead of orthogonally optimizing all weights (ablation) | RegO · EVDA Avg | EER (%): 7.63 -> 6.34 | Helped |
| Add Ebbinghaus Forgetting Mechanism (EFM) (ablation) | RegO · CLEAR Exp_10 | Accuracy (%): 94.85 -> 94.75 | No clear effect |
| Add Importance Region Localization (IRL) (ablation) | RegO · CLEAR Exp_10 | Accuracy (%): 92.83 -> 94.75 | Helped |
| Add Region-Adaptive Optimization (RAO) instead of orthogonally optimizing all weights (ablation) | RegO · CLEAR Exp_10 | Accuracy (%): 90.91 -> 94.75 | Helped |
| Remove Ebbinghaus Forgetting Mechanism (EFM) (ablation) | RegO · CLEAR Exp_1 | Accuracy (%): 91.92 -> 92.93 | Helped |
Surprising findings
- Removing the Ebbinghaus forgetting mechanism improved results on CLEAR image recognition, contrary to its benefit on audio deepfake detection.
- Finetune-Exp1 results on CLEAR show accuracy remains similar early on, suggesting retained old knowledge may interfere with later tasks.
Limitations
- Relies on a fixed threshold alpha and forgetting threshold gamma selected by hyperparameter study.
- Gains over baselines on EVDA are modest on several experiences (e.g., Exp1 second-best behind EWC), and image results use frozen features with a small classifier.
Experimental setup
Datasets: EVDA (FMFCC, In the Wild, ADD 2022, ASVspoof2015, ASVspoof2019, ASVspoof2021, FoR, HAD) CLEAR
Benchmarks: EVDA CLEAR
Metrics: EER accuracy
Baselines: Finetune EWC GDumb GEM CWRStar SI OWM RAWM RWM Replay-All
Builds on: Elastic Weight Consolidation Orthogonal Weight Modification Radian Weight Modification Ebbinghaus forgetting curve
Closest papers by meaning
- What to Remember: Self-Adaptive Continual Learning for Audio Deepfake Detection (AAAI 2024 · similarity 0.91)
- Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection (NeurIPS 2026 · similarity 0.88)
- I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.86)
- Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning (CVPR 2025 · similarity 0.85)
- Multi-modal Deepfake Detection via Multi-task Audio-Visual Prompt Learning (AAAI 2025 · similarity 0.85)
- AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (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.