PIM-FFL: Evaluating Participation Index Metrics for Fairness Enhancement in Federated Learning

Federated learning (FL) is a technique that allows multiple devices to train a machine learning model without sharing their raw data, this preserves user privacy. However, traditional FL often suffers from problems like nodes failing to participate in training due to resource limitations. This negatively impacts the model's performance and fairness, making it harder for all devices to benefit equally. Ensuring fairness in FL is challenging due to the significant differences in clients' data quality, data size, and contribution levels. A notable issue arises when prime contributing clients do not receive fair rewards while underperforming still benefits. This disparity leads to an inequity causing complexity among contributors. As a result, clients may feel unappreciated, leading to a decrease in their interest in participating which ultimately compromises the complete reliability and effectiveness of the system. To address these challenges, a client clustering method is presented based on their participation index, which ensures that prime participants receive better model updates rather than underperformers which makes the system more reliable. Additionally, the Double Deep Q-Networks (DDQN) method is employed, as it determines the best actions through learning from past experiences and updating Q-values, which dynamically update participation count and reduce the model biases. Our experiments show that this clustering method substantially enhances fairness in Federated Learning, fostering a more efficient and cooperative learning environment that outperforms traditional methods like FedAvg and FedProx.