Comprehensive Validation of Bridge Module and EBM Loss for One-Class Audio Deepfake Detection[link]
Seung-Bin Kim, Hyun-Seo Shin, Jungwoo Heo, Chan-Yeong Lim, Kyo-Won Koo, Jisoo Son, Souhwan Jung, Ha-Jin Yu
abstract
In audio deepfake detection, a one-class perspective is emerging as a promising route to improved generalization. Following this view, we previously proposed a framework that employs an enhanced bona fide modeling (EBM) loss. The EBM loss is designed to encourage bona fide utterances to form a compact and coherent normal region in the embedding space, while explicitly pushing deepfake utterances away from this region. In other words, it systematically reshapes the decision boundary around the distribution of bona fide speech so that a one-class decision criterion can generalize to unseen attack types. However, the prior study did not systematically examine training-related factors (e.g., augmentation policy and input utterance length) or model dependence across diverse frontends and backbones. Accordingly, this paper conducts a focused validation along four axes: 1) data augmentation; 2) hyperparameters of loss function; 3) train–test utterance length; and 4) portability across several frontends and backbones. Through experiments conducted on the ASVspoof 2019 LA, 2021 LA, and 2021 DF datasets, we confirm that the EBM framework consistently improves performance across diverse architecture combinations and provide concise guidelines for robust deployment under distribution shift.
Comprehensive Validation of Bridge Module and EBM Loss for One-Class Audio Deepfake Detection[link]
Seung-Bin Kim, Hyun-Seo Shin, Jungwoo Heo, Chan-Yeong Lim, Kyo-Won Koo, Jisoo Son, Souhwan Jung, Ha-Jin Yu
abstract
In audio deepfake detection, a one-class perspective is emerging as a promising route to improved generalization. Following this view, we previously proposed a framework that employs an enhanced bona fide modeling (EBM) loss. The EBM loss is designed to encourage bona fide utterances to form a compact and coherent normal region in the embedding space, while explicitly pushing deepfake utterances away from this region. In other words, it systematically reshapes the decision boundary around the distribution of bona fide speech so that a one-class decision criterion can generalize to unseen attack types. However, the prior study did not systematically examine training-related factors (e.g., augmentation policy and input utterance length) or model dependence across diverse frontends and backbones. Accordingly, this paper conducts a focused validation along four axes: 1) data augmentation; 2) hyperparameters of loss function; 3) train–test utterance length; and 4) portability across several frontends and backbones. Through experiments conducted on the ASVspoof 2019 LA, 2021 LA, and 2021 DF datasets, we confirm that the EBM framework consistently improves performance across diverse architecture combinations and provide concise guidelines for robust deployment under distribution shift.