MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202611065759 A) filed by Noida Institute Of Engineering And Technology Niet on May 25, 2026, for Probe-Based Defense System For Securing Federated Large Language Model Training Against Malicious Clients.
Inventors include Harsh Chauhan; and Sneha Rai.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: The present invention relates to a probe-based defense system (100) for securing federated large language model training against malicious clients. The system comprises a LoRA-Probe module (104) extracting incremental B-matrix differences from low-rank adaptation modules (103) in the first transformer layer of each client, constructing L2-normalized feature vectors from concatenated weight deltas, and applying a pre-trained linear classifier (114) to compute maliciousness probability scores. A safety defense module (105) transforms these scores into aggregation security factors through step-level (106), client-level (107), and shadow-level (108) defense mechanisms. A shadow low-rank adaptation branch (115) provides distribution-drift-resistant detection signals. A security- weighted aggregation engine (109) normalizes client contributions using sample count and security factor, while a security-gated controller (110) prevents aggregation during rounds dominated by malicious clients. The system provides robust defense with minimal computational overhead for privacy-sensitive distributed computing environments.
Disclaimer: Curated by HT Syndication.
