A multi-level and iterative feature engineering framework for deepfake audio detection

Abstract

This study introduces FakeSleuthNeXt, a lightweight, fully interpretable, handcrafted feature engineering framework for deepfake audio detection. The method relies exclusively on manually designed features without any neural network training. It combines a histogram-driven binary pattern extractor applied to 7-level discrete wavelet transform sub-bands with 15 statistical descriptors per level. An iterative ensemble feature selection strategy fusing INCA, IChi2, and IReliefF algorithms produces compact and highly discriminative representations. Evaluated on six challenging and diverse deepfake audio datasets (over 72,000 segments) using only kNN and Cubic SVM classifiers, the framework achieves accuracies ranging from 89.20% to 99.21%, EER values from 0.97% to 10.85%, and min-tDCF of 0.124 (ASVspoof 2019 LA) and 0.298 (ASVspoof 2021). These results match or surpass many recent deep learning systems while offering significantly lower computational cost, full transparency, and straightforward deployment on resource-constrained devices, making FakeSleuthNeXt particularly suitable for forensic applications. FakeSleuthNeXt provides a fast, transparent, and highly resource-efficient solution, making it particularly well-suited for forensic applications and deployment on resource-constrained devices.