10th WorldS4 2026

Authors - Makomborero Murwira, Dane Brown
Abstract - The proliferation of realistic synthetic speech poses significant threats to information integrity and public safety through financial fraud and misinformation campaigns. While traditional countermeasures based on Gaussian Mixture Models have proven effective against earlygeneration deepfakes, they struggle to generalise to sophisticated attacks produced by contemporary neural synthesis. This paper investigates enhanced audio deepfake detection through Light Convolutional Neural Networks (LCNNs) combined with Linear Frequency Cepstral Coefficients (LFCCs) and advanced training strategies. This study systematically evaluates the impact of margin-based loss functions (Cosface and A-Softmax) and FreqAugment data augmentation on model robustness and generalisation capability. Validated on the ASVspoof 2019 Logical Access dataset, the optimised LCNN model achieves an Equal Error Rate (EER) of 5.40% on the evaluation set containing thirteen unseen attack types, representing a 33% relative improvement over the baseline LFCCGMM countermeasure (8.09% EER). The combination of Cosface loss with FreqAugment demonstrates superior performance compared to ASoftmax configurations, reducing false negatives substantially. Per-attack analysis reveals robust performance across diverse spoofing techniques, though vulnerabilities to advanced neural waveform manipulation methods remain. The proposed framework provides a practical, deployable solution for audio deepfake detection in real-world security-critical applications.