A Deep-Reinforcement Learning Approach to Mitigate Vishing Attacks and Data Theft in Voice over Internet Protocol Calls

Voice over Internet Protocol (VoIP) technology has transformed global communication by providing cost efficiency, scalability, and integration with digital platforms. Nevertheless, its widespread adoption has increased exposure to cyber threats, including vishing attacks, call spoofing, and data breaches. Vishing, a social engineering technique designed to manipulate users into disclosing confidential information, has become particularly prevalent, underscoring the need for intelligent and adaptive defense mechanisms. This study developed a Deep- Reinforcement Learning (DRL) framework to detect and mitigate VoIP threats. The framework integrates deep neural networks with reinforcement learning, enabling adaptive decisionmaking and proactive security responses. Development comprised objective formulation, data set preparation, feature engineering and iterative training using historical VoIP-to-cellular attack datasets and simulated datasets generated within a virtual testbed. The DRL model achieved 90.0% accuracy in historical data (precision 88.7%, recall 91.3%, F1-score 90.0%) and 89.24% accuracy in simulated data (precision 87.5%, recall 90.6%, F1-score 89.0%), with false-positive rates below 6% and an average detection latency of 2.1 seconds. Comparative benchmarking against Support Vector Machines, Random Forest, and Convolutional Neural Networks confirmed superior performance, highlighting the scalability, robustness, and practical significance of the framework in securing VoIP communications across evolving environments.