MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641075005 A) filed by Vellore Institute Of Technology on June 17, 2026, for System And Method For Privacy-Preserving Federated Learning Of Biometric Data.
Inventors include Ragavan K; and Lekkala Dathri.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: Embodiments herein provide a privacy-preserving federated learning system (100) for biometric data including a central server (104) configured to aggregate model parameters and two or more client devices (102). Each client device (102) includes a memory (106) and processors (108) configured to determine, before local model training, a bounded gradient-based input modification to biometric input data using an iterative Projected Gradient Descent process constrained within a predefined epsilon bound, train a machine learning model using both unmodified and corresponding modified biometric input data within a same training cycle, and transmit model parameters to the central server (104) while excluding transmission of raw biometric data, modified biometric data, and intermediate representations. The processors (108) vary the epsilon bound across training rounds after a warm-up phase based on model convergence, wherein gradient information is altered to reduce susceptibility to gradient-based reconstruction while maintaining classification performance. FIG. 1
Disclaimer: Curated by HT Syndication.
