Neural Forensic Analysis for Privacy and Integrity Protection in Biometric Authentication Systems

This paper proposes a digital forensics-driven approach to ensure the integrity and privacy preservation of biometric systems in consumer electronics. Specifically, we develop a neural network-based forensic model to detect and analyze disguised speech attacks, which pose a significant threat to biometric authentication. The model leverages Multi-Layer Perceptron (MLP) architecture to identify speaker gender based on key acoustic parameters such as formant center frequency, bandwidth, and sound intensity. To enhance forensic accuracy, we employ L-BFGS optimization during model training. Experimental validation is conducted using SoundTouch-modified speech samples, simulating real-world biometric spoofing attempts. We further analyze the impact of network structure and activation functions on detection performance, as well as the model's adaptability to various electronic disguise techniques. Results demonstrate that the proposed MLP-based forensic framework effectively differentiates between genuine and electronically disguised speech, providing a robust solution for biometric security in consumer electronics.