Executive Summary : | Federated learning systems (FLS) are set to become the future of AI due to the decentralized availability of data and growing concerns about privacy. FLS is a server-client model for training machine learning algorithms, where clients train the model locally using local data and only communicate the trained model to the server. The server aggregates the models trained by these clients and proposes a new starting model to all clients, ensuring data privacy and respect for all available data. Bias in machine learning is an emerging topic, especially when applied to societal applications like job applications, healthcare, and education. This can arise from imbalanced data or existing societal biases, such as fewer job opportunities for female candidates due to fear of leaving the job after getting married or having children. The new paradigm of FLS addresses data collection issues but also has constraints, as each client may have data pertaining to its own demographics. The project aims to address two questions in FLS: 1) Does the existing centralized fair algorithm extend to FLS? and 2) How clients can be incentivized to train their local model and reveal their model accurately to the server. The project plans to identify clients that perform well on accuracy and fairness using multi-armed bandit techniques. Combinatorial MABs with non-availability constraints will be built to select a subset of clients, even if some clients may not be present. Incentives mechanisms for clients to contribute to FLS and data pre-processing techniques on the client side will be developed to improve accuracy and fairness even when their data is heterogeneous. |