近年來在深度偽造語音偵測系統中，基於 wav2vec 2.0 預訓練模型做為前端特徵提 取器的方法在許多資料集上取得不錯的表現， 然而在 deepfake audio 的鑑識裡面， 往往 不只需要正確的答案， 也需要一個解釋去說明答案的依據， 而這類自監督學習所產生 的潛在表徵本身缺乏語義可解讀性，加上處理的後端模型亦是黑箱的情況下，導致傳 統的事後歸因方法難以追溯回原始語音。此外， wav2vec 2.0 模型以原始波形作為唯一 輸入，得到的解釋也僅能反應某段時間的重要性，並不能擴展到頻率域上做後續的分 析。 因此我們提出模型無關後設代理模型框架，在不犧牲這類模型優秀的表現外，藉 由訓練另一個代理的解釋模型，將 mel spectrogram(梅爾頻譜圖)表徵對齊至 wav2vec 的 模型表徵， 以達到取得解釋的效果。透過預訓練過的 Audio Spectrogram Transformer (AST) 作為 mel spectrogram 的特徵提取器搭配交叉注意力投影器逼近黑盒的表徵，並將 交叉注意力與 AST 原本的注意力一起展開熱圖視覺化原本黑盒模型決策關注的時頻區 域。在五個 ASVspoof 等基準測試集上全面評估，證實代理模型於決策層（ EER 一致 性）與表徵層（ MSE 對齊），並透過忠實度測試驗證代理模型是否正確顯示原黑箱模 型的決策。 本研究的貢獻有三： 1. 提出一個與模型無關且不犧牲模型分類精確度的代理模型 框架； 2. 打破原本受限於原始音訊輸入的限制，擴展解釋至頻率域上面； 3. 透過忠實 度分析確保後設解釋與代表模型的決策過程。;Recent advances in deepfake audio detection have demonstrated strong performance using pretrained wav2vec 2.0 models as front-end feature extractors. However， audio forensics demands not only accurate classification but also interpretable justifications. The latent representations produced by self-supervised models lack semantic transparency， and with black-box back-end classifiers， conventional post-hoc attribution methods cannot reliably trace decisions back to the original speech signal. Furthermore， since wav2vec 2.0 operates on raw waveforms， its explanations are confined to the temporal domain and cannot be extended to frequency-domain analysis. To address this， we propose a model-agnostic surrogate explanation framework that preserves detection performance while providing interpretable visual explanations. A pretrained Audio Spectrogram Transformer (AST) serves as the mel-spectrogram feature extractor， paired with a cross-attention projector to approximate the black-box model′s representation space. Cross-attention weights are combined with AST′s internal attention via attention rollout to generate time-frequency heatmaps visualizing the black-box model′s decision-critical regions. Comprehensive evaluations on five ASVspoof benchmark datasets confirm surrogate fidelity at both the decision level (EER consistency) and the representation level (MSE alignment)， with faithfulness tests further validating that the surrogate′s explanations genuinely reflect the original black-box model′s decision process. This work makes three contributions: (1) a model-agnostic surrogate framework that maintains classification accuracy; (2) extension of explanations beyond the temporal domain into the frequency domain; and (3) faithfulness analysis ensuring that the surrogate′s explanations reliably reflect the black-box model′s decision-making process.
