Research map / Label-efficient and robust learning / Recognition and long-tail

Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AAFI (input-dependent routing over a heterogeneous compression pool of Whisper features) and Mel-BAP (PGD adversarial perturbations restricted to a 4–8 kHz mel band) for generalized audio deepfake detection, reaching 1.51% average EER with Whisper-small.

Problem

Audio deepfake detectors generalize poorly across heterogeneous spoofing artifacts, unseen attacks, and languages, and standard augmentation degrades clean accuracy (robustness-accuracy trade-off).

Approach

Adds a noisy top-k router that fuses multiple linear-compressed latent resolutions of frozen Whisper encoder features, and applies PGD adversarial training restricted to a fixed high-frequency mel band of the log-Mel input, with the Whisper decoder used as a token-generating classifier.

Key results

• Whisper-small + AAFI + Mel-BAP achieves 1.51% average EER across five benchmarks, with 1.87% on ADD 2022 Track 1 vs 9.96% for Resemble-Detect-3B-Omni.
• Ablation on SpeechFake BD-UT: Whisper-base EER falls from 1.19% (w/o Mel-BAP) to 0.91% (Mel-BAP only) and 0.53% (AAFI+Mel-BAP).
• Global augmentation raises Whisper-base clean EER on BD-UT from 1.19% to 2.33%, while Mel-BAP lowers it to 0.91%.

Surprising findings

• Global augmentation with the same distortion types substantially degrades clean EER, supporting the claimed robustness-accuracy trade-off.
• Top-2 AAFI routing performs worse than the baseline (5.73% vs 3.52% average EER on validation groups).

Limitations

• Routing capacity sensitivity (top-k choice affects performance).
• Cross-system comparison uses heterogeneous and undisclosed training protocols from Speech-DF-Arena, so it is only contextual.
• Mel-BAP adds training cost proportional to PGD iterations; real-time scalability in non-stationary environments is unresolved.

Experimental setup

Datasets: SpeechFake BD SpeechFake MD CodecFake ADD 2022 FakeOrReal In-the-Wild ASVspoof 2019 ASVspoof 2021 LA NoiseX-92

Benchmarks: ADD 2022 Track 1 FakeOrReal In-the-Wild ASVspoof 2019 ASVspoof 2021 LA SpeechFake BD-UT SpeechFake MD unseen languages

Metrics: EER average EER

Baselines: Resemble-Detect-3B-Omni DLMSL-SpeakSure-v0.1 MoLEx DF_Arena_1B_V_1 XLSR + SLS BiCrossMamba-ST Whisper-base w/o AAFI+Mel-BAP Global Aug

Builds on: Whisper PGD adversarial training Frequency-Selective Adversarial Training (F-SAT) mixture of experts

Closest papers by meaning

• I Can Hear You: Selective Robust Training for Deepfake Audio Detection (ICLR 2025 · similarity 0.89)
• Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning (CVPR 2025 · similarity 0.88)
• Region-Based Optimization in Continual Learning for Audio Deepfake Detection (AAAI 2025 · similarity 0.88)
• AVFF: Audio-Visual Feature Fusion for Video Deepfake Detection (CVPR 2024 · similarity 0.86)
• What to Remember: Self-Adaptive Continual Learning for Audio Deepfake Detection (AAAI 2024 · similarity 0.86)
• SLIM: Style-Linguistics Mismatch Model for Generalized Audio Deepfake Detection (NeurIPS 2024 · similarity 0.85)