Swarm Learning Empowered Federated Deep Learning for Seamless Smartphone-Based Activity Recognition

In the landscape of smartphone-based human activity recognition (S-HAR), adopting Federated Deep Learning (FDL) introduces challenges, notably in communication inefficiencies and data confidentiality. These issues stem from the requisite submission of learning model parameters by multiple clients to FDL's global model. To surmount these challenges, the innovative Swarm Learning (SL) paradigm emerges as a solution, presenting a modular approach that fuses distributed computing with blockchain-based coordination. This amalgamation eliminates the dependence on a centralized infrastructure. This study introduces an avant-garde Swarm-Federated Deep Learning framework (SHAR-SFDL) that seamlessly incorporates SL into the FDL framework. SHAR-SFDL orchestrates the collaboration of smartphone users in creating local SL models through blockchain-enabled synergy. The aggregation of these local models into a global FDL model across diverse SL groups is achieved through a groundbreaking method involving model credibility prediction and weight comparison. Notably, the proposed SHAR-SFDL framework showcases a substantial advancement in model performance and a remarkable reduction in edge-to-global communication overhead. Regarding performance evaluation, the proposed model outperformed the other state-of-the-art techniques regarding true and false positive rates across different group densities. Specifically, the TP rates for SHAR-SFDL were 0.891 (High), 0.945 (Medium), and 0.969 (Low), while the corresponding FP rates were 0.035 (High), 0.009 (Medium), and 0.015 (Low).