CoRELoop: Parameter-Efficient Controlled Recurrent Refinement for Audio Deepfake Detection

Authors: Kunyu Feng, Yuxiang Wang, Li Wang, Wan Lin, Zhizheng Wu

Published: 2026-09-17 07:28:40+00:00

Comment: 5 pages, 2 figures, 3 tables

AI Summary

CoReLoop proposes a parameter-efficient method for improving audio deepfake detection by introducing controlled recurrent refinement into an already-trained, frozen self-supervised learning (SSL)-based detector. It leverages lightweight refinement modules and low-rank adapters to enable additional passes without modifying the original large model parameters or requiring new training data. This approach significantly reduces the equal error rate across diverse unseen attacks while maintaining parameter efficiency.

Abstract

Generalizing to unseen attacks remains challenging for audio deepfake detectors, and collecting training data covering all potential attacks is impractical. We explore recurrent refinement in an already-trained SSL-based detector without additional data or changes to its original parameters. However, directly recycling encoder outputs as inputs degrades detection in our diagnostic. We propose CoReLoop, which makes this reuse effective by adapting recurrent inputs to the frozen encoder, controlling state updates, and aligning refined outputs with the frozen classifier. By training only lightweight refinement modules and loop-specific low-rank adapters on the original data, CoReLoop enables additional refinement while preserving the detector's original first-pass prediction. On 14 cross-domain test sets, the 24-layer model reduces pooled equal error rate (EER) from 4.85% to 3.74% with two passes, with approximately 10M trainable parameters out of 598M. To selectively apply this refinement, an optional halting head chooses the depth for each utterance, achieving 3.73% pooled EER with an average of 1.18 passes.

Key findings

CoReLoop reduces the pooled Equal Error Rate (EER) from 4.85% to 3.74% with two passes on 14 cross-domain test sets, using only approximately 10M trainable parameters out of 598M. Adaptive inference with a halting head achieves a pooled EER of 3.73% with an average of 1.18 passes, demonstrating significant performance gains with substantial inference cost reduction compared to fixed multi-pass approaches.

Approach

The approach integrates recurrent refinement into a frozen SSL-based audio deepfake detector. It uses a "LoopBridge" to adapt previous encoder outputs as recurrent inputs, employs loop-specific Low-Rank Adaptation (LoRA) for parameter-efficient updates within the frozen encoder, and controls state updates with an "UpdateGate." An optional halting head dynamically determines the optimal number of refinement passes for each utterance.

Datasets

ASVspoof 2019 LA, ADD 2023, DFADD, AISHELL-3, ITW, ASVspoof, FoR, CF, ADD, DFADD, LSV, SONAR (14 public speech datasets, specific names not fully listed but implied by references).

Model(s)

Wav2vec 2.0 BERT (referred to as W2V-BERT, with 3, 6, 12, and 24 Transformer layers)

Author countries

China

← Previous