Research map / Label-efficient and robust learning / OOD and anomaly detection
Learning Tight Rejection Boundaries without Negatives for Strict One-Class Audio Deepfake Detection
Yuze Zhao, Kuiyuan Zhang, Zhongyun Hua, Yushu Zhang, Qing Liao, Wei Jiang
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 centroid-anchored boundary shaping (OBJcabs) with off-manifold probes (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 33.83 -> 13.44 | Improved |
| add benign-invariance objective (OBJbinv) (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 16.43 -> 13.44 | Improved |
| use centroid-referenced margin instead of BCE on shifted views (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 21.56 -> 13.44 | Improved |
| use centroid-referenced margin instead of InfoNCE on shifted views (ablation) | CA-SOADD · ASVSpoof 2021 LA/DF and ASVSpoof-5 | EER (%): LA 15.97 -> 7.30; ASVSpoof-5 16.95 -> 13.44 | Improved |
| add benign invariance loss instead of using benign transforms as augmentation only (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 19.15 -> 13.44 | Improved |
| swap XLSR backbone for WavLM or HuBERT (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 13.44 -> 16.38 (WavLM), 19.90 (HuBERT) | Hurt |
| add online standardization with bona fide statistics (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 17.84 -> 13.44 | Improved |
| add learned layer selection instead of uniform pooling (ablation) | CA-SOADD · ASVSpoof-2021 LA | EER (%): 7.03 -> 7.30 | Mixed |
| add learned time weighting instead of uniform pooling (ablation) | CA-SOADD · ASVSpoof-5 test, whole | EER (%): 14.88 -> 13.44 | Improved |
| enable OBJbinv with language-conditioned centroids on MLAAD (ablation) | CA-SOADD · MLAAD | EER | Hurt |
| apply strict real-only CA-SOADD versus spoof-supervised AASIST on zero-shot transfer | ASVSpoof-2021 to ASVSpoof-5 · ASVSpoof-2021 to ASVSpoof-5 transfer | EER (%): 39.9 -> 14.7 | Improved |
Closest papers by meaning
- Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection (NeurIPS 2026 · similarity 0.87)
- Region-Based Optimization in Continual Learning for Audio Deepfake Detection (AAAI 2025 · similarity 0.85)
- Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning (CVPR 2025 · similarity 0.85)
- I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.84)
- SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake Detection (NeurIPS 2024 · similarity 0.83)
- What to Remember: Self-Adaptive Continual Learning for Audio Deepfake Detection (AAAI 2024 · similarity 0.83)
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.
