Communication Signal Modulation Recognition Based On Federated Learning

Modulation recognition is an important research topic in the field of signal intelligent processing,and its goal is to identify the modulation of the signal to be examined,so as to achieve effective acquisition of signal content in non-collaborative communication.Deep learning-based modulation recognition has been widely used because it utilizes the powerful feature extraction ability of deep learning and avoids the disadvantages of traditional methods,such as tedious process and poor stability to a certain extent.However,there are some challenging problems in the current deep learning-based modulation recognition,such as the excessive number of model parameters and computational effort,the insufficient sample size or sample dimension of labeled signals leading to difficulties in training a high-quality model,and the privacy leakage problem caused by directly aggregating multiple signal samples to enrich the sample size.To address these issues,the work conducted in this thesis is as follows:（1）A lightweight neural network-based modulation recognition model TSB-CLDNN is proposed.Converting the original I/Q signal into instantaneous amplitude and phase through data pre-processing,and using multiple features of the signal as model input;using Bottleneck structure convolution,bi-directional long and short term memory networks,channel attention and multiple channel merging operations to form the model.Experimental results on the Radio ML 2016.10a dataset shows that TSB-CLDNN performs better in the recognition accuracy,number of parameters and computational volume relative to the five classical models;the accuracy is improved by 0.1%,the number of parameters is reduced by 64.8%,and the computational volume is reduced by 48.6%relative to MCLDNN,which has the highest accuracy among the classical models.（2）Fed MR,a federated learning-based modulation recognition method is proposed,which uses the Fed Avg aggregation algorithm for weighted average aggregation of model parameters to jointly train modulation recognition models without aggregating local signal data among the participants.The experimental results shows that the global model trained by Fed MR has improved the recognition accuracy,recall and precision of various signals compared with the local model trained by each participant using only local signal data,indicating that Fed MR can effectively alleviate the problem of data resource barriers restricting the direct aggregation of multiple signal samples,which hinders the further improvement of modulation recognition model recognition performance.（3）A federated learning modulation recognition method DP-Fed MR,which combines Fed Prox aggregation algorithm and differential privacy,is proposed.DP-Fed MR improves on Fed MR by using Fed Prox instead of Fed Avg as the aggregation algorithm to alleviate the problem of poor model performance caused by data heterogeneity in federated learning training by adding a proximal term to the loss function;adding differential privacy noise to the parameters of the interaction model to resist attack methods against model parameters such as model inversion attacks,and enhancing the degree of privacy protection for the local signal data of each participant.The experimental results shows that the Fed Prox aggregation algorithm improves the recognition accuracy of all types of signals by an average of 1.6%in the case of data heterogeneity;DP-Fed MR loses only 0.5%accuracy when the privacy budget is 1 and the relaxation term is 10^-3,indicating that the improvement of DP-Fed MR is effective.（4）Fed MR is implemented in a distributed environment using the Fed ML open source platform for federated Learning,using two Ubuntu system hosts as participants in the Federated Learning task,and using the public cloud server provided by Fed ML official for model parameter aggregation.The detailed steps of Fed MR implementation using Fed ML platform are presented,and the visualization results after the training is completed are shown.