Federated Learning For Image Classification based on Deep Learning

Abstract

Federated Learning has been emerged as a promising for modern Machine Learning techniques. Classical manner of operating in a centralize dataset come up against critical privacy issues. Beside that real data reacted with real user’s behavior is beneficial to tasks which involve model to be trained on practical data. For example, language model can be leveraged by playing on user data emitted while they text for speech recognition or next word prediction tasks. We could also utilize images on end devices to improve image classification models. Two current state-of-the-art methods when dealing with federated system are FedAvg and FedProx. While FedAvg proposed a heuristic algorithm that is quite robust about independent and identically distributed distribution (IID), the latter further upgrade upon the local loss setting for stability with respect to the non-IID distribution. There are two main nature challenges within the task as indicated in FedProx work: system heterogeneity and statistical heterogeneity. One more dif-ficulty: the lack of a systematic hyperparameters tuning as well as model selection approach. FedAvg and FedProx mostly work with canonical datasets and their synthesis variants like MNIST, CIFAR-10. In this work, we employ the Federated Learning approaches to unusual dataset to observe the capabilities of generalizing when handling domain-specific tasks. Con-cretely, we adopt FedAvg and FedProx on: (1) a brain tumor dataset with 3064 512×512 T1-weight images and (2) a VNPlant-200 dataset which includes 20,000 images of 200 unique medicinal plants. Following the work in FedAvg and FedProx, two algorithms are applied with a careful hyperparameter tuning and inspect the effect of federated setting on the decentralized environment. The work empirically demonstrates the impact of federated learning on distinct domains. In addition, the experiments provide a heuristic scheme for hyperparameter control-ling in other similar tasks or data, in this case, distributed model training and brain tumor or medicinal plant datasets.