深度學習模型近年來在多項語音相關任務中展現了卓越的效能，但其對對抗性攻擊的脆弱 性也引發了廣泛關注，特別是在安全關鍵的應用場景中更顯嚴重。為此，本研究提出一種防 禦框架，結合高斯雜訊與語音增強模組，用以重建遭受對抗性擾動的輸入語音。我們的方法 能有效消除對抗性擾動，同時保留語音的感知品質與語意可解釋性。 在 SC09、VCTK 及 Qualcomm Keyword Spotting（QKWS）三個常用語音資料集上的實驗顯 示，該方法在面對多種威脅模型下皆展現出優異的強健性，包括具不同限制範數與步長的 PGD 攻擊、採用期望變換（EOT）策略的自適應白盒攻擊，以及如 FakeBob 所代表的查詢 式黑盒攻擊。此外，與現有先進防禦方法相比，我們的方法在推論速度與模型大小方面皆具 明顯優勢，極適合應用於即時性與資源受限之系統中。 實驗結果充分證明了本研究在強化語音系統強健性方面的有效性與實用性。;Deep learning models have achieved state-of-the-art performance across various speech-related tasks; however, their vulnerability to adversarial attacks poses a significant challenge, particularly in safety-critical applications. In this work, we propose a defense framework that leverages a speech enhancement module with Gaussian noise to purify adversarial inputs. Our method effectively removes adversarial perturbations while preserving speech quality and semantic interpretability. Extensive experiments on SC09, VCTK, and QKWS datasets demonstrate that our approach achieves superior robustness under both white-box and black box threat models, including PGD with different steps, norm constraints, and EOT budgets, as well as query-based attack such as FakeBob. Moreover, our method outperforms prior defenses in terms of inference speed and model compactness, making it suitable for real-time, resource constrained deployments. These results highlight the effectiveness and efficiency of our approach as a practical defense mechanism for robust speech systems.
