MUMBAI, India, April 17 -- Intellectual Property India has published a patent application (202631044782 A) filed by Sanjib Kumar Rout, Bhubaneswar, Odisha, on April 8, for 'scra: reliability-controlled federated learning system with selection-governed participation and adaptive aggregation.'
Inventor(s) include Kishan Kanta Barik; Sujit Pattnaik; Swastik Prasad Behera; Sradha Ram; Achirangshu Patra; and Rajat Kumar Samantaray.
The application for the patent was published on April 17, under issue no. 16/2026.
According to the abstract released by the Intellectual Property India: "A federated learning system is disclosed that trains a shared machine learning model across a plurality of distributed client devices while preserving data locality. The system introduces a closed-loop control framework wherein client behaviour is continuously evaluated through a timeevolving reliability state maintained at a central server. The reliability state is updated each federated learning round using composite quality indicators derived from client model updates, encompassing model divergence, temporal stability, gradient characteristics, and local training loss. A participant orchestration module leverages the reliability state along with directional cosine alignment scores to selectively admit or exclude client devices from participation in each training round, thereby preventing degraded updates from influencing global model convergence. A reliability-aware scheduling module determines the local training allocation for each participating client as a function of its reliability and stability, ensuring that consistent contributors are assigned proportionally greater training effort. A trust-adaptive aggregation module combines client updates into a refined global model using composite weights that incorporate dataset size, reliability exponent scaling, directional alignment, and stability scores derived from curvature-based analysis of successive update trajectories. An optional differential privacy module enables noise injection and gradient clipping for privacy-preserving deployments. Byzantine bound tracking records firstexclusion round indices and computes per-round exclusion fractions for empirical robustness analysis. By integrating behavioural evaluation, temporal tracking, controlled selection, adaptive scheduling, and constrained aggregation within a unified reliability-driven framework, the system achieves improved convergence stability and reduced influence of noisy or adversarial client contributions in heterogeneous distributed learning environments."
Disclaimer: Curated by HT Syndication.
