An Integrated ViT–DQN Framework for Jamming Detection and Adaptive Mitigation in Wireless Communications

Radio interference is a problem that impacts the quality of communications and reliability of wireless communications. In the light of the above, the aim of this research is to develop an automatic radio interference detection and classification system based on deep learning and a "Vision Transformer" (ViT) model. This was done by converting the received “IQ” signals into color coded spectral windows that were 224×224 pixels in size. These windows are then passed to the model “ViT-B/16” where each one is classified as either a valid signal or one of three types of interference: harmonic interference, broadband interference, and intermittent interference. A balanced window set of 20,000 windows was constructed, with 14,000 windows used for training, 3,000 windows used for validation and 3,000 windows used for testing. The system was trained in two steps: first, the base model was trained using the same weights in the classification head, and then the last four layers of the Transformer encoder and the classification head were fine-tuned. The model was fine-tuned based on the results of an independent test set, which found that the accuracy was 93.70% and the Macro-F1 score was 0.9362. The model performed well in recognizing both tonal interference and broadband interference. The confusion matrix demonstrated that the main challenge for the system is intermittent interference as most of the errors were due to the difficulty in distinguishing between normal and intermittent interference. These results confirm the effectiveness of using the Vision Transformer (ViT) model to analyze the spectral representation of IQ signals for the purpose of detecting and classifying radio interference.