Frequency-Selective Adversarial Attack in WSN Detected using Deep Learning Algorithms

Adversarial attacks in Wireless Sensor Networks (WSNs), such as jamming and spoofing, threaten the reliability and security of communication systems. This paper presents a novel adversarial detection framework based on sub-band energy levels using a deep learning (DL) architecture with multi-task learning (MTL). The detection system, implemented on a Jetson Nano platform, processes incoming modulation signals susceptible to attack. The signals are first preprocessed using the Transverse Dyadic Wavelet Transform (TyDWT), which decomposes them into frequency sub-bands to eliminate noise and extract critical features. Subsequently, the Rational Dilation Wavelet Transform (RADWT) and Tunable Q-factor Wavelet Transform (TQWT) are applied to preserve and adaptively analyze the signal's full frequency spectrum, focusing on identifying abnormal oscillations linked to adversarial interference. The processed signal features are then analyzed by a DL model that classifies normal and adversarial signals. By utilizing multi-task learning, the model efficiently handles multiple tasks, such as anomaly detection, signal classification, and attack identification. The proposed system demonstrates a robust capability to detect adversarial activities and ensure secure communication in WSNs.