Optimization Design Of Multimodal Learning In Wireless Networks

Artificial intelligence is the key enabling technology and the main application scenario of 6G in the future.The multi-dimensional wireless data are generated by user behavior,network environment and wireless devices.These data in wireless networks provide a data basis for network intelligence.This multi-dimensional information acquisition framework gradually highlights the multimodal characteristics of wireless data.In order to accurately describe the characteristics of devices or users and mine the relevance of multimodal data,multimodal learning needs to be introduced into wireless network.Multimodal learning,which is one of the important methods to deal with multimodal data,aims to establish a model that can correlate multimodal data.Using multimodal learning to analyze the data of wireless network can fully explore the potential and advantages of high dimension and large volume of wireless data,and greatly improve the efficiency and accuracy of intelligent processing and control of wireless network.However,the multimodal data and computing power of wireless network are scattered among users and edge devices,which requires a large number of communications and interactions between nodes to realize distributed cooperation and joint multimodal learning processing.Due to the large number of devices participating in multimodal learning,and the limited communication resources of wireless network,it is difficult to ensure the reliability of multimodal learning,which leads to the user privacy disclosure and low communication efficiency in multimodal learning of wireless network.This thesis studies the optimization of multimodal learning in wireless network.The specific research work can be summarized as follows:This thesis studies the privacy protection method of wireless transmission security for multimodal learning.Firstly,a privacy protection method of multimodal learning wireless transmission based on adding noise into data is designed.Then,the model accuracy performance of multimodal learning in wireless network is analyzed and the expression of the upper bound of the model accuracy loss is derived.The effectiveness of the proposed method in privacy protection of multimodal data transmission is illustrated.Secondly,a joint optimization problem for the model accuracy loss and data security risk is proposed,in order to realize the compromise between the model accuracy performance of multimodal learning and the data security of wireless transmission,and further a joint optimization algorithm to optimize the noise power,user scheduling and sample selection strategies is designed.Finally,the simulation experiments verified that the optimization scheme can greatly reduce the security risk of wireless transmission at the cost of little model accuracy loss.This thesis studies the communication overhead optimization method of wireless transmission for multimodal learning.Firstly,a wireless communication overhead optimization method of multimodal learning based on data compression is designed.Similarly,the model accuracy performance of multimodal learning in wireless network is analyzed.The compromise relationship between model compression efficiency and model accuracy loss is clarified.Secondly,in order to keep the balance between the accuracy performance of multimodal learning and wireless transmission communication overhead,the compression matrix and user scheduling strategy are jointly optimized,and then an optimization method based on mixed integer nonlinear programming is proposed.Finally,the simulation and experiment results show that the proposed optimization scheme can reduce the communication overhead to a great extent at the cost,and has little influence on the model accuracy loss of multimodal learning.