<![CDATA[Deepfake audio has become a serious threat to digital security due to its ability to generate highly realistic synthetic speech for malicious purposes such as impersonation, fraud, and social engineering. Despite the rapid advancement of deepfake generation technologies, audio-based detection remains relatively underexplored compared to image and video domains. Moreover, existing approaches largely rely on handcrafted features and conventional machine learning methods, which are often insufficient to capture complex temporal and spectral dependencies and often exhibit limited generalization to unseen deepfake generation techniques. This study proposes a novel framework that leverages log-Mel spectrogram representations to capture informative time–frequency characteristics of audio signals. The proposed model integrates a Hybrid Audio Spectrogram Transformer to model global dependencies through self-attention and is further enhanced by a Dual-Attention Feature Fusion mechanism to improve discriminative capability at both spatial and channel levels. In addition, Bayesian Optimization is employed to efficiently optimize hyperparameters, leading to improved model performance, faster convergence, and better generalization. Experimental results demonstrate that the proposed model achieves an accuracy of 93.60% and an AUC of 0.9724, indicating excellent discriminative performance. The model also achieves a high recall in detecting deepfake audio, highlighting its effectiveness in minimizing false negatives and improving robustness in real-world scenarios. These findings demonstrate that the proposed framework provides an effective and reliable solution for deepfake audio detection and offers significant potential for deployment in practical cybersecurity and digital audio forensic applications.]]>
Copyrights © 2026
