Design And Implementation Of Federated Learning Platform Based On Android

Traditional machine learning generally processes data on a single machine or cluster,trains the model,and then issues the model in the cloud.The accuracy of the algorithm is highly dependent on the training and analysis of massive data.Although large-scale collection of data greatly improves the performance of machine learning algorithms,it also brings greater challenges to the protection of personal privacy data,especially on mobile terminals.Personal privacy data security issues are further highlighted.Based on the above background,this thesis proposes a Horizontal Federation Learning strategy based on the Android system.Based on this Horizontal Federation Learning strategy,this thesis designs and implements a distributed machine learning system based on mobile terminal.The system trains the model on each Android terminal based on the data provided by the user to generate a personalized model.Through the server integration of the personalized model uploaded by the Android terminal,the shared model is finally obtained.This thesis elaborates on the theoretical basis,design ideas,implementation schemes,testing and analysis of the system design.The main tasks of this analysis are as follows:(1)Provide Android-based model inference,model training,and model update.This analysis transplants TensorFlow.js to the Android platform and implements model inference and model training through TensorFlow.js.The Android terminal downloads the latest cloud shared model,and improves and trains the model based on local user data on the Android terminal.The system extracts the improved model as a small update file and uploads the update file to the cloud.(2)Provide Android terminal scheduling management and training data management.The system proposes a detailed scheduling strategy based on the status of the device to ensure that mobile phone user experience will not be degraded by Federal Learning.Training data management is responsible for deleting and adding data under appropriate conditions.(3)The server designs a Federated Learning model sharing strategy and issues the model.The server receives the model update file,weights the average model parameters according to the model update file,and designs a loss function to evaluate the model performance.(4)This thesis conducts the function and performance of the distributed machine learning system based on mobile terminal.and analyzes the test results in detail.