Intellectual Property India Publishes Patent Application for System And Method For Secure Tensor Layout Scrambling In Edge Cloud Split Inference

MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202631085005 A) filed by Brainware University on July 10, 2026, for System And Method For Secure Tensor Layout Scrambling In Edge Cloud Split Inference.

Inventors include Dr. Saurabh Pal; Mr. Partha Shankar Nayak; Mr. Shuvrajit Nath; Mr. Saurav Bhaumik; Mr. Prem Kumar; and Mr. Jitesh Prasad Khatick.

The application for the patent was published on July 17, 2026, under issue no. 29/2026.

Abstract: The present disclosure relates to a system and method for secure tensor layout scrambling in edge-cloud split inference. An edge- native security apparatus (102) receives input data from a raw sensor/data source (104) and uses neural execution hardware (106) with a split-layer execution unit (108) to generate an intermediate feature-map tensor (110), which is stored in a volatile tensor buffer (112). A cryptographic permutation controller (114), including a cryptographic secret storage unit (116), a nonce generation unit (118), a permutation seed derivation unit (120), and a pseudorandom permutation map generator (122), generates a nonce-dependent pseudorandom permutation map. An address-remapping memory-transfer controller (124), optionally including a DMA engine (124A) and a scatter-gather descriptor table (126), rearranges tensor index positions to generate a scrambled tensor stored in a scrambled tensor buffer (128) while leaving numerical activation values unchanged. A secure serialization unit (130) serializes the scrambled tensor, and a metadata authentication unit (132) generates protected reconstruction metadata with an authentication tag (166). A network interface (134) transmits the serialized scrambled tensor and protected reconstruction metadata through a communication network (136) to a cloud-side inference system (138). The cloud-side inference system (138) verifies the metadata using a metadata verification unit (140), accesses cryptographic material through a cloud-side cryptographic secret or session key unit (141), regenerates an inverse permutation using an inverse permutation controller (142), restores tensor ordering within a protected tensor reconstruction environment (144), and completes inference using a remaining neural network execution unit (146) to produce an inference output (148).

Disclaimer: Curated by HT Syndication.