Harnessing Deep Learning Architectures for Interference Detection and Classification in ISAC Systems

Interference is a major challenge in integrated sensing and communication (ISAC) systems, and accurate interference classification is critical for ensuring reliable system design. We generate four types of ISAC signals, three of which incorporate self-interference, mutual interference, and clutter interference. We employ three deep learning architectures, including deep neural network (DNN), convolutional neural network (CNN), and long short-term memory (LSTM) models, for interference detection and classification. We compare their performance under different signal-to-noise ratio (SNR) conditions, focusing on both classification accuracy and computational complexity. Simulation results indicate that the CNN outperforms both the DNN and LSTM models in terms of accurate interference classification, albeit at a higher computational complexity. In contrast, the DNN exhibits the lowest complexity but compromises classification accuracy, particularly under low SNR conditions. The LSTM model provides the most balanced performance, effectively balancing classification accuracy and computational complexity.