WIDE NEURAL NETWORKS APPROACH FOR DETECTING AND MITIGATING STOCHASTIC DECEPTION ATTACKS IN CYBERSECURITY

This study addresses the critical challenge of detecting and mitigating stochastic deceptive attacks in cyber environments through a novel hybrid approach combining behavioural analysis and Wide Neural Networks (WNN). The proposed solution introduces a new integrated data model that combines attack and attacker characteristics, enhanced through generative adversarial techniques for data augmentation. The core innovation of the proposed approach involves a WNN architecture optimized with a bio-inspired trophallaxis regularization approach to prevent overfitting during classification. Experimental evaluation of 2-, 4-, and 6-layer configurations revealed significant findings: Without trophallaxis, deeper networks showed declining performance (6-layer: 49% validation accuracy). With trophallaxis, the 4-layer WNN achieved optimal balance (89% training, 59% validation accuracy), while the 2-layer model overfit (87% training, 50% validation) and the 6-layer showed diminishing returns (89% training, 54% validation).System implementation demonstrated 89% attack detection success against sophisticated threats (IP rotation, content obfuscation, redirection) with <13% false positives. Real-time countermeasures applied for the mitigation of the threat including traffic throttling and quarantine protocols proved effective in operational testing. These results establish the 4-layer WNN with trophallaxis as an optimal solution that offers superior accuracy-generalization trade-offs for real-world cybersecurity applications. The study advances deception attack mitigation through its unique integration of behavioural modelling, bio-inspired regularization, and practical system implementation.