A poisoning attack method for electromagnetic signal modulation type identification based on a pure label

Abstract

A poisoning attack method for electromagnetic signal modulation type identification based on a pure label comprises the following steps: firstly, building an electromagnetic signal classification model, and enabling an electromagnetic signal data set to predict and identify a modulation type with high precision; adopting a pure label poisoning attack method, generating a poisoning sample by searching a sample similar to the feature space of the target sample; finally, using a forward and backward splitting iteration method for optimization, and after an optimal poisoning sample is generated, using transfer learning for retraining the classification network; finally, enabling the classification model to generate misclassification, so that a poisoning attack effect is achieved. According to the poisoning attack method for electromagnetic signal modulation type identification based on a pure label, the poisoning attack can be carried out under the condition that training data does not need to be known and only a training model and parameters thereof need to be known, and a poisoning defense measure measured based on the training data is invalid.

Description

Technical field [0001] The present invention belongs to the field safety study depth, relates to an attack poisoning pure electromagnetic tag signal modulation type recognition. Background technique [0002] With the continuous development of artificial intelligence, depth research study has been widely used in image recognition, voice recognition, radio modulation type classification and other fields. Although the neural network has a high accuracy rate, but very fragile. Therefore, by attacks on deep learning model, and then we found the holes in the model was extremely important. Poisoning attacks occurred in the model training phase, the attacker through the back door triggers embedded in deep learning model. Further, the input sample poisoning in the testing phase, the goalkeeper is triggered to attack. [0003] Poisoning attacks against deep learning model includes two aspects: First: data poisoning, the attacker first get poisoned poisoned data to the training part of the ...

AI Technical Summary

Problems solved by technology

Second: Model poisoning, the attacker directs the model structure and model parameters, the modified model can achieve user requirements with a very high accuracy rate, but it will show poisoning behavior for specific target samples

This method not only requires the attacker to have a good understanding of the original data, but also the performance of the entire classifier will be affected during the poisoning training process, which is difficult to achieve in many cases

Images