Intellectual Property India Publishes Patent Application for System And Method For Cryptographically Keyed Localized Tensor Precision Degradation In Split- Inference Networks

MUMBAI, India, July 30 -- Intellectual Property India has published a patent application (202631085004 A) filed by Brainware University on July 10, 2026, for System And Method For Cryptographically Keyed Localized Tensor Precision Degradation In Split- Inference Networks.

Inventors include Dr. Ranadip Kundu; Dr. Suparna Panchanan; Mr. Prabir Kumar Das; Mr. Dulal Adak; and Mr. Mainul Hasan.

The application for the patent was published on July 24, 2026, under issue no. 30/2026.

Abstract: The present invention relates to a system and method for cryptographically keyed localized tensor precision degradation in split-inference networks. A split-inference system (100) includes an edge computing device (110), a communication network (160), and a remote inference server (170). The edge computing device (110) includes a mixed-precision neural processing unit (120) configured to generate an intermediate activation tensor, a cryptographic schedule generator (130) configured to generate a deterministic pseudorandom precision map (136) from a secret seed and a session-dependent value, a hardware truncation controller (140) configured to remove one or more lower-order bit-planes from localized tensor regions according to the deterministic pseudorandom precision map (136), and a tensor packing buffer (150) configured to store a variable-precision representation of the intermediate activation tensor. The variable-precision representation is transmitted through a DMA / network interface path (154) and the communication network (160) to the remote inference server (170). The remote inference server (170) regenerates or obtains a corresponding deterministic pseudorandom precision map (136) and processes the received variable-precision representation using a degradation-aware neural network component (174). The invention may reduce reconstruction fidelity of transmitted intermediate tensors for unauthorized parties while supporting split-inference processing by the authorized remote inference server (170).

Disclaimer: Curated by HT Syndication.