AI-Driven Anomaly Detection in Wireless Sensor Networks

Abstract

The uses of WSNs can be applied to smart infrastructure, industrial monitoring and cyber-physical systems among other mission-critical uses. The topology of distribution and dynamic traffic patterns along with high power demand renders the WSNs extremely susceptible to any form of anomaly due to node failures, environmental disturbance and even by malicious attacks. Non-flexible methods of detection of anomalies include statistical as well as threshold based methods which in non-stationary network environment give high false alarm. The current research was intended to develop a flexible and energy efficient AI-based anomaly detection system which will have the capability of accurately identifying abnormal behavior with less computational overheads on WSNs. The presented framework integrates two approaches that are complimentary. Firstly, a hybrid EWMALSTM/GRU model will be learned and capable to detect both gradual and abrupt sensor data anomalies with short-term statistical variance and the long-term temporal contexts. Second, Deep Reinforcement Learning (DRL) agent is an agent that optimizes the accuracy of anomaly detection and energy consumption by co-opting to learn the optimum detection thresholds and policies towards responses in the course of its ongoing interaction with the network environment. The benchmark WSN data in terms of traffic loads and attack conditions were experimented over long periods of time. The proposed framework achieved a detection rate of 96.9 and F1-score of 95.8 and reduced false positive rate of 2.6, which was over 10.7 per cent more precise than individual LSTM, GRU and fixed statistical models. Moreover, adaptive policy that was implemented based on DRA reduced redundant transmissions by 34.2 and this contributed significantly to the long life of the network. The results show that the hybrid AI model that is proposed in this research is scalable, adaptive, and energy-efficient in anomaly detection of the next generation WSN.