Medicinal Plants Identification Using Federated Deep Learning

Abstract

Over the years, scientists have discovered bioactive chemicals in many of the plants that have been traditionally utilized as medicinal medicines. However, identifying plant species based on their physical characteristics can be difficult, and misidentification can have severe consequences, such as the use of the incorrect plant as a medicine. With the advent of machine learning techniques such as deep learning and federated learning, it is now possible to develop automated systems for the precise image-based classification of medicinal plants. Nevertheless, medicinal plant classification using deep learning techniques typically requires a large amount of data, which can be challenging to acquire and manage due to privacy concerns, data ownership, and geographic reasons. Federated learning provides a solution to this issue by enabling the training of a shared model on multiple devices without requiring centralized data storage. In this work, we assess and optimize the federared learning framework using two federated learning approaches, FedAvg and FedProx, and four state-of-the-art deep learning networks for the job of categorizing medicinal plants by distributing the original training set into two forms, IID and Non-IID. Ultimately, the accuracy of the optimal federated learning system is improved by 5.65% and 14.84% over the baseline on IID data and Non-IID data, respectively. Furthermore, the study brings up a new difficult arena for the task of classifying medicinal plants using Non-IID training data.