MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202621036769 A) filed by Persistent Systems on March 26, 2026, for System And Method For Dynamic Privacy–utility Optimization In Obfuscated Large Language Model Inference Processing.
Inventor includes Mr. Nitish Shrivastava.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: A system [500] and method for dynamically optimizing the privacy-utility tradeoff in obfuscated large language model processing through continuous quality assessment and feedback-driven strategy adaptation. The system [500] implements a comprehensive feedback loop that continuously measures output quality across multiple dimensions, identifies which privacy parameters most significantly impact quality through statistical analysis, and proposes and implements parameter adjustments to move toward an optimal pareto frontier in privacy-quality space. A quality scoring model [530] compares obfuscated responses against reference responses from a held-out evaluation set. The system [500] employs a multi-dimensional quality score comprising four independent dimensions: technical correctness assessment, semantic fidelity, factual consistency, and reasoning coherence. A strategy impact analyzer [540] utilizes analysis of variance and two-way interaction effect modeling to identify significant parameters. A parameter adjustment engine [550] implements a conservative step-size policy with automatic rollback. A privacy constraint enforcer [570] maintains a minimum privacy score of 0.90.
Disclaimer: Curated by HT Syndication.
