Research On Federated Learning For Mobile Image Recognition Algorithm

The rapid progress of artificial intelligence research has promoted the implementation of many intelligent applications,such as machine translation,speech recognition,face recognition,intelligent recommendation,etc.,which greatly facilitates people’s production and life.However,it also raises the issue of people’s privacy data leakage,especially in the case of image recognition on mobile devices.In this paper,an image recognition framework based on federated learning is proposed to protect the data privacy of users while improving performance of image recognition algorithms in mobile devices.Specifically,through joint modeling in multiple mobile devices,only model parameters are exchanged while user data are kept on local devices.which realizes privacy protection on the basis of solving data islands.Then to improve model performance,this paper further designs attention-based federated aggregation and metric-based constraint.Moreover,to deal with the common few-shot learning scenarios,this paper introduces meta-learning into proposed framework.To summarize,the main research points and innovations of this paper are as follows.1.A federated learning algorithm based on Hierarchical Attention is proposed for image recognition on mobile devices.The algorithm utilizes a light deep neural network model and an efficient parameter transmission mechanism to fit mobile devices,meanwhile,a federated aggregation method based on Hierarchical Attention mechanism is designed to enhance the effect of federated aggregation.Experimental results illustrate that the algorithm can be effectively used for image recognition tasks of mobile devices.2.A distance constraint-based federated learning algorithm of mobile devices is proposed.The algorithm compliments original image recognition loss with a new distance-constrained loss function to optimize intra-class and inter-class distances of samples,which is applied to the aforementioned federation learning framework to further improve the model performance.Experimental results show that the algorithm can directly improve local model training process,and further improve the performance of aggregated model.3.A federated meta-learning algorithm based on metric learning is proposed.The algorithm combines federated learning and meta-learning methods to improve generalization ability and class expansibility in the small sample learning scenario.Experiments illustrate that the algorithm can be effectively applied to the small sample learning scenario under the federal learning framework.