A Deep Learning Approach with Neuroevolution and Feature Selection for Voice Authentication Spoofing Detection

• Grant number: 25/10220-5
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: November 01, 2025
• End date: October 31, 2028
• Field of knowledge: Engineering - Electrical Engineering
• Principal Investigator: Rodrigo Capobianco Guido
• Grantee: Monique Simplicio Viana
• Host Institution: Instituto de Biociências, Letras e Ciências Exatas (IBILCE). Universidade Estadual Paulista (UNESP). Campus de São José do Rio Preto. São José do Rio Preto , SP, Brazil

Abstract

Among the biometric modalities employed in authentication systems, human voice stands out due to its ease of acquisition and its strong technical and scientific foundation in the field of signal processing. However, the increasing sophistication of spoofing attacks, particularly replay attacks, in which recordings of authorized individuals' voices are reused by impostors to bypass security systems, has exposed structural vulnerabilities in traditional voice verification systems. In this context, initiatives such as the Automatic Speaker Verification and Spoofing Countermeasures Challenge (ASVspoof) have been encouraging the development of more effective countermeasures, through the provision of standardized datasets and comparative evaluation protocols. This project proposes the development of a methodology for fraud detection in voice authentication, structured around three fundamental components: deep learning, feature selection, and neuroevolution applied to the modeling and optimization of deep neural networks. In this proposal, neuroevolution is explored as a strategic mechanism for the automatic generation of high-performance architectures, enabling the dynamic adaptation of network topology, as well as the optimization of relevant hyperparameters, through the application of evolutionary algorithms, with emphasis on genetic algorithms and their hybridizations with specialized local search operators and genetic improvement strategies. The performance of the proposed methodology will be validated using benchmark datasets provided by the ASVspoof challenges, and the results will be disseminated in peer-reviewed scientific journals and high-impact conferences, as well as made publicly available in specialized repositories, ensuring the reproducibility of experiments and contributing to the advancement of the field.