Research map / Large language models / Language, safety and interpretability
FoeGlass: Simple In-Context Learning Is Enough for Red Teaming Audio Deepfake Detectors
Sepehr Dehdashtian, Jacob H Seidman, Vishnu Boddeti, Gaurav Bharaj
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 |
|---|---|---|---|
| use FoeGlass cold start instead of unconditional sampling | LLM attacker on xTTS-v2 vs VIT-VoxCelebSpoof-ConstantQ · FNR | false negative rate (%): 2.24 -> 80.72 | Improved |
| use warm start (three examples) instead of cold start | FoeGlass on Kokoro-82M vs VIT-VoxCelebSpoof-MelSpectrogram · FNR | false negative rate (%): 7.52 -> 39.72 | Improved |
| use warm start instead of cold start | FoeGlass on xTTS-v2 vs AST-ASVspoof5 · FNR | false negative rate (%): 4.86 -> 3.97 | No clear effect |
| add diversity feedback (ablation) | FoeGlass on xTTS-v2 vs AST-VoxCelebSpoof · FNR | false negative rate (%): 7.10 -> 63.30 | Improved |
| add diversity feedback (ablation) | FoeGlass on xTTS-v2 vs VIT-VoxCelebSpoof-ConstantQ · FNR | false negative rate (%): 86.02 -> 96.29 | Improved |
| add diversity feedback (ablation) | FoeGlass on xTTS-v2 vs VIT-ASVspoof5-ConstantQ · FNR | false negative rate (%): 96.33 -> 93.76 | Hurt |
| add chain-of-thought to attacker context (ablation) | FoeGlass on xTTS-v2 vs AST-VoxCelebSpoof · FNR | false negative rate (%): 58.39 -> 63.30 | Improved |
| add chain-of-thought to attacker context (ablation) | FoeGlass on xTTS-v2 vs VIT-VoxCelebSpoof-ConstantQ · FNR | false negative rate (%): 81.66 -> 96.29 | Improved |
| add realness feedback only (no diversity) vs unconditional sampling (ablation) | LLM attacker · FNR | false negative rate (%) | Improved |
| fine-tune detector on FoeGlass data instead of unconditional data | RawNetLite · held-out VITS | reported % (labeled accuracy): 29.6 -> 8.2 | Improved |
| fine-tune detector on FoeGlass data from RawNetLite attacks | AASIST · held-out VITS | reported % (labeled accuracy): 15.2 -> 0.2 | Improved |
| transfer FoeGlass attacks to other detectors | FoeGlass cold start · 8 ADDs | false negative rate | Improved |
Closest papers by meaning
- Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection (NeurIPS 2026 · similarity 0.89)
- I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.87)
- SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake Detection (NeurIPS 2024 · similarity 0.85)
- ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models (NeurIPS 2025 · similarity 0.84)
- Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning (CVPR 2025 · similarity 0.84)
- What to Remember: Self-Adaptive Continual Learning for Audio Deepfake Detection (AAAI 2024 · similarity 0.84)
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.
