DESIGNING ATTACK-RESISTANT AND INTELLIGENT TRUST MANAGEMENT USING DEEP NEURAL NETWORKS

Abstract

The rapid expansion of IoT and cyber-physical systems has exposed critical weaknesses in trust management, making networks vulnerable to sophisticated adversarial threats. Conventional trust models often lack the adaptability and robustness required in dynamic environments. To address this, we propose an intelligent trust management framework powered by deep neural networks that not only detects attacks with high accuracy but also demonstrates resilience against adversarial manipulation. The approach integrates advanced pre-processing, feature selection, and balancing techniques, followed by classification using LSTM, GRU, and a hybrid LSTM–GRU architecture, where the hybrid consistently surpassed standalone models in both binary and multiclass tasks. For the CAN dataset, convolution-based models, including a lightweight VGG16, provided efficient and reliable detection of complex attack patterns. Beyond detection, the system was rigorously tested under adversarial conditions such as evasion, poisoning, and inference attacks, and showed notable resistance through strategies like adversarial training. Overall, this dual capability of accurate attack detection and strong attack resistance highlights the scalability, reliability, and security of the proposed system, making it well-suited for modern IoT trust management.