Abstract:
With the growing awareness regarding user privacy and the demand for transparency, traditional distributed machine learning algorithms are becoming taboo that uses the user’s data without any restriction or privacy. The corporation uses this paradigm to harvest user data to gain more insights into user interaction with their application to provide a better experience. However, the end-user has no control over the private data used by these companies. This necessitates the adoption of a decentralized paradigm in which the data is safe with the user. Federated Learning (FL) is a distributed learning paradigm that can learn a global model from decentralized data without exchanging sensitive data across the users. Through our work, we intend to study the application of Federated Learning in the domain of Computer Vision for image classification. Through our thesis, we aim to understand the real-world scenario where different commercial image sources can collaborate in a Federated setting to perform the image classification task with privacy preservation. Most previous research works applied the Federated Learning algorithm on a single dataset distributed among the clients in an IID or non- IID manner. This is not close to a real-world scenario where different clients may have different data distributions due to domain shifts. To address this, we propose our own dataset derived from 8 different commercial sources to understand the application of Federated Learning in real-world scenarios and understand how different commercial sources can collaborate in a federated setting. Also, another challenge in Federated Learning is the convergence issue when the data distribution is different among the clients, which may increase the communication cost between the clients and the central server and leads to suboptimal model performance. For this part, we specifically worked on the Domain shift issue. To tackle this, we worked on proposing two novel methods, namely Fed-Cyclic and Fed-Star. In the final part of our work, we worked on the problem of class-label imbalance among different clients and explored different techniques to mitigate the issue and create robust models that can effectively learn from non-IID data.