Research map / Label-efficient and robust learning / Recognition and long-tail
SONAR: Spectral‑Contrastive Audio Residuals for Generalizable Deepfake Detection
Ido Nitzan Hidekel, Gal Lifshitz, Khen Cohen, Dan Raviv
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 SRM high-pass noise branch and JS alignment loss to dual XLSR encoders (ablation) | SONAR-Full · In-the-Wild | EER (%): 8.91 -> 6.00 | Improved |
| add JS alignment loss (ablation) | SONAR-Full · ITW | EER (%): 8.50 -> 6.00 | Improved |
| add JS alignment loss (ablation) | SONAR-Full · ASVspoof 2021 LA | EER (%): 2.90 -> 1.55 | Improved |
| use learnable SRM filters instead of fixed ones (ablation) | SONAR-Full · In-the-Wild | EER (%): 8.20 -> 6.00 | Improved |
| add the SRM rich feature extractor (ablation) | SONAR-Full · In-the-Wild | EER (%): 8.44 -> 6.00 | Improved |
| use separate (unshared) XLSR encoder weights for content and noise branches (ablation) | SONAR-Full · In-the-Wild | EER (%): 8.40 -> 6.00 | Improved |
| increase number of SRM filters from 1 to 30 (ablation) | SONAR-Full · In-the-Wild | EER (%): 8.00 (M=1) -> 7.40 (M=10) -> 6.00 (M=30) | Improved |
| raise JS loss weight lambda from 0.5 to 1 (ablation) | SONAR-Full · In-the-Wild | EER (%): 7.91 (0.5) -> 7.02 (0.8) -> 6.00 (1) | Improved |
| add SONAR NFE and JS alignment to pre-trained XLSR-Mamba | XLSR-Mamba · In-the-Wild | EER (%): 6.71 -> 5.43 | Improved |
| add SONAR to XLSR-Mamba | XLSR-Mamba · ASVspoof 2021 LA | EER (%): 0.93 -> 1.20 | Hurt |
| add SONAR alignment to baseline encoders | XLSR-AASIST and XLSR-Mamba · In-the-Wild high-pass >4 kHz probe | EER (%): 32.51 -> 26.57; 29.77 -> 26.08 | Improved |
| replace AASIST head with two-layer MLP (ablation) | SONAR · In-the-Wild | EER (%): 6.00 -> 6.98 | Hurt |
Closest papers by meaning
- Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection (NeurIPS 2026 · similarity 0.87)
- AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (CVPR 2024 · similarity 0.86)
- Region-Based Optimization in Continual Learning for Audio Deepfake Detection (AAAI 2025 · similarity 0.86)
- 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 0.85)
- DeepASA: An Object-Oriented Multi-Purpose Network for Auditory Scene Analysis (NeurIPS 2025 · 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.
