MUMBAI, India, Aug. 10 -- Intellectual Property India has published a patent application (202641090933 A) filed by Vellore Institute Of Technology on July 27, 2026, for Decentralized Federated Learning System With Gradient Deduplication Using Dual Cryptographic Fingerprinting.
Inventors include Varalakshmi M; and Ankit Subedi.
The application for the patent was published on July 31, 2026, under issue no. 31/2026.
Abstract: ABSTRACT DECENTRALIZED FEDERATED LEARNING SYSTEM WITH GRADIENT DEDUPLICATION USING DUAL CRYPTOGRAPHIC FINGERPRINTING A method for decentralized federated learning with gradient deduplication is provided. Edge devices (112) within an edge tier (110) train local machine learning models to generate gradient updates. A gradient fragmentation module (130) with a layer-aware partitioner (132) partitions the gradient updates into fixed-size gradient fragments aligned with neural network layer boundaries. A fingerprint module (134) generates a primary cryptographic fingerprint using a BLAKE3 hash generator (136) and a secondary cryptographic fingerprint using a SHA-256 hash generator (138) for each gradient fragment. A smart contract layer (124) of a distributed ledger tier (120) executes a deduplication process comprising searching a fragment index (140), verifying candidate matches, storing unique fragments in a fragment repository (142), and generating reference identifiers via a reference identifier generator (144). A reconstruction module (150) reconstructs complete gradient updates that are mathematically equivalent to the original gradient updates. The reconstructed gradient updates are aggregated to produce an updated global model stored in distributed ledger storage (126). (FIG. 1)
Disclaimer: Curated by HT Syndication.
