Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Alethia: a Foundational Encoder for Voice Deepfakes
Yi Zhu, Brahmi Dwivedi, Jayaram Raghuram, Surya Koppisetti
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 generative flow-matching objective (ablation) | Alethia-Base · SDD-Eval-50 (all) | EER: +0.5 EER when removed | Improved |
| add generative flow-matching objective (ablation) | Alethia-Base · SDD diffusion and flow-matching attacks | EER: +5.6 EER when removed | Improved |
| add generative flow-matching objective (ablation) | Alethia-Base · SDD other generators | EER: -0.1 EER when removed | No clear effect |
| add bottleneck architecture without masking or generative objective (ablation) | Alethia-Base · SDD vs WavLM-Large | EER: 0.7% EER improvement, 0.49% accuracy gain | Improved |
| continue pretraining Wav2vec2-Base on 1k hours real+fake speech with masked token prediction | W2V-base · SDD 50 datasets | EER: +0.80 | Hurt |
| pretrain Wav2vec2-Base from scratch on real+fake speech | W2V-base · SDD 50 datasets | EER: +1.20 | Hurt |
| pretrain HuBERT-Base from scratch on real+fake speech | HuBERT-base · SDD 50 datasets | EER: -0.30 | No clear effect |
| continue pretraining HuBERT-Base on real+fake speech | HuBERT-base · SDD 50 datasets | EER: +0.25 | Hurt |
| raise CNN mask ratio from 10% to 20% (ablation) | Alethia-Base · SDD 30 datasets | EER: 16.2 -> 19.2 | Hurt |
| raise CNN mask ratio from 10% to 30% (ablation) | Alethia-Base · SDD 30 datasets | EER: 16.2 -> 19.3 | Hurt |
| use zero mask instead of learnable CNN mask (ablation) | Alethia-Base · SDD 30 datasets | EER: 16.2 -> 17.4 | Hurt |
| use Gaussian noise mask instead of learnable CNN mask (ablation) | Alethia-Base · SDD 30 datasets | EER: 16.2 -> 18.9 | Hurt |
Closest papers by meaning
- 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 0.87)
- 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)
- Multi-View Collaborative Learning Network for Speech Deepfake Detection (AAAI 2025 · similarity 0.85)
- Phoneme-Level Feature Discrepancies: A Key to Detecting Sophisticated Speech Deepfakes (AAAI 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.
