Robust signal identification method and device for patch attack positioning and clearing

By constructing a robust signal recognition method for locating and eliminating patch attacks, and utilizing time-frequency graph transformation and unsupervised network detection to repair adversarial patches, the problem of deep learning models being vulnerable to patch attacks in signal modulation recognition is solved, achieving high accuracy and robust recognition in multi-patch scenarios.

CN121010877APending Publication Date: 2025-11-25XIDIAN UNIV
0 Cites 0 Cited by

Patent Information

Application Number
CN202510980869.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-16
Publication Date
2025-11-25

AI Technical Summary

Technical Problem

Deep learning models are vulnerable to patch attacks in the field of signal modulation recognition. Existing defense methods have poor scalability in multi-patch attack scenarios and cannot guarantee the accuracy and robustness of the classifier.

Method used

By constructing a robust signal recognition method for patch attack localization and removal, and using pseudo-Wigner-Villi distribution to transform time-series signals into time-frequency maps, an unsupervised patch attack localization network and image reconstruction network are trained. Combined with convolutional neural networks, detection and repair are performed, and adversarial patch samples are generated to test the defense capability.

Benefits of technology

It effectively detects and locates adversarial patches, reduces their impact on classification results, and ensures the accuracy and robustness of signal modulation type identification, especially outperforming random smoothing methods in multi-patch attack scenarios.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a robust signal identification method and device for patch attack positioning and clearing. The method comprises the following steps: obtaining a training set and a test set; dividing the time-frequency graph of the training set into a plurality of image blocks for training an unsupervised patch attack positioning network; dividing the time-frequency graph of the training set into a plurality of image blocks, and performing random masking on the image blocks; constructing a time-frequency graph classification network model by using a convolutional neural network, and training a classification network by using the time-frequency graph of the training set until a network loss function converges; detecting the image block through the patch attack positioning network, and performing mask processing on a potential attacked image to obtain a masked image block; the masked image blocks and the unmasked image blocks are sent to an image reconstruction network for restoration, and a restored image is obtained; sending the restored image into a classification network to identify a modulation type; a single patch attack sample and a multi-patch attack sample are generated by using an anti-patch generation method to test the attack resistance of the method.
Need to check novelty before this filing date? Find Prior Art