MUMBAI, India, June 26 -- Intellectual Property India has published a patent application (202641072767 A) filed by Vellore Institute Of Technology on June 11, 2026, for System And Method For Robust Federated Learning Under Client Dropouts Using Partial Update Salvaging.
Inventors include R Adaline Suji; Shirsh Harishnakar Bhakta; and Sarthak Dey.
The application for the patent was published on June 19, 2026, under issue no. 25/2026.
Abstract: The present disclosure proposes a system (100) and method for robust federated learning under client dropout conditions using partial update salvaging, weighted partial aggregation, checkpoint-based preservation of intermediate training states, and transmission of partially trained model parameters in decentralized edge- computing environments operating under intermittent connectivity conditions. The system (100) comprises plurality of client devices (101), an aggregation server (122), and a computing device (102) configured to collectively perform decentralized federated model training in distributed edge-computing environments operating under intermittent connectivity conditions. The proposed system (100) is compatible with standard federated learning architectures without requiring layer-wise architectural modification of machine learning models. The proposed system (100) improves utilization of interrupted distributed client-side computation during federated model updating operations. The proposed system (100) is capable of maintaining model updating operations under intermittent connectivity conditions without relying on stale model updates, substitute client updates, or cached historical model parameters.
Disclaimer: Curated by HT Syndication.
