Binarization in DeepFake Audio Detection: A Comparative Study and Performance Analysis

Abstract

DeepFake audio, generated through advanced AI techniques, poses significant risks such as fraud, misinformation, and identity theft. As the quality of synthetic audio improves, detecting such fakes has become increasingly challenging. Traditional detection methods struggle to keep pace as AI-generated voices replicate speech patterns, tone, and pitch convincingly. While computationally intensive large-scale models can help detect DeepFakes generated by AI, their resource requirements make them impractical for deployment on mobile devices as well as on resource-constrained devices. This paper proposes a lightweight yet effective approach using binarized neural networks (BNNs) and further enhancements using additional dense layers and stacked modeling to overcome these challenges. We conduct a comprehensive performance analysis of the network and compare it with various machine learning and neural network methods to evaluate the tradeoff between detection accuracy and computational efficiency as an effect of binarization and precision loss in feature embeddings. © 2025 IEEE.