AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Mechanistic analysis of attention and representations · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use FoeGlass cold start instead of unconditional samplingLLM attacker on xTTS-v2 vs VIT-VoxCelebSpoof-ConstantQ · FNRfalse negative rate (%): 2.24 -> 80.72Improved
use warm start (three examples) instead of cold startFoeGlass on Kokoro-82M vs VIT-VoxCelebSpoof-MelSpectrogram · FNRfalse negative rate (%): 7.52 -> 39.72Improved
use warm start instead of cold startFoeGlass on xTTS-v2 vs AST-ASVspoof5 · FNRfalse negative rate (%): 4.86 -> 3.97No clear effect
add diversity feedback (ablation)FoeGlass on xTTS-v2 vs AST-VoxCelebSpoof · FNRfalse negative rate (%): 7.10 -> 63.30Improved
add diversity feedback (ablation)FoeGlass on xTTS-v2 vs VIT-VoxCelebSpoof-ConstantQ · FNRfalse negative rate (%): 86.02 -> 96.29Improved
add diversity feedback (ablation)FoeGlass on xTTS-v2 vs VIT-ASVspoof5-ConstantQ · FNRfalse negative rate (%): 96.33 -> 93.76Hurt
add chain-of-thought to attacker context (ablation)FoeGlass on xTTS-v2 vs AST-VoxCelebSpoof · FNRfalse negative rate (%): 58.39 -> 63.30Improved
add chain-of-thought to attacker context (ablation)FoeGlass on xTTS-v2 vs VIT-VoxCelebSpoof-ConstantQ · FNRfalse negative rate (%): 81.66 -> 96.29Improved
add realness feedback only (no diversity) vs unconditional sampling (ablation)LLM attacker · FNRfalse negative rate (%)Improved
fine-tune detector on FoeGlass data instead of unconditional dataRawNetLite · held-out VITSreported % (labeled accuracy): 29.6 -> 8.2Improved
fine-tune detector on FoeGlass data from RawNetLite attacksAASIST · held-out VITSreported % (labeled accuracy): 15.2 -> 0.2Improved
transfer FoeGlass attacks to other detectorsFoeGlass cold start · 8 ADDsfalse negative rateImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy