Efficient Deep-Learning-Based Decision Fusion for Adversarial Sensor Networks With Byzantine Attacks

This paper proposes a deep learning-based approach to resilient decision fusion in adversarial sensor networks under Byzantine attacks. A unified probabilistic framework is first introduced to capture diverse adversarial settings, including varying ratios of malicious nodes, synchronized and unsynchronized attacks, Markovian system states, unbalanced priors, and adaptive attacking strategies. Building on this formulation, a deep neural network is trained on a globally constructed dataset to infer system states directly from received node reports over an observation window, eliminating the need for scenario-specific models or explicit parameter tuning. Extensive simulations demonstrate the model’s robustness across all tested scenarios, achieving up to 99.97% accuracy, bit error rates below 0.001, and fast inference suitable for real-time sensing applications. The results show that deep learning can serve as a powerful and scalable alternative to traditional fusion rules in adversarial multi-sensor environments.