Intellectual Property India Publishes Patent Application for Softpuf: A Software Based Physically Unclonable Function Framework For Device Binding Of Machine Learning Models On Commodity Edge Hardware

MUMBAI, India, June 26 -- Intellectual Property India has published a patent application (202641071698 A) filed by Ms. N. Uma; Ms. S. Kavishree; Dr. A. Indumathi; Ms. R. Saktheeswari; Ms. M. Sugacini; Mr. V. Ranjith; Mr. S. Chandrasekar; Mr. P. V. Dhanesh Sivadeep; and Mr. S. Gokhula Anand on June 09, 2026, for Softpuf: A Software Based Physically Unclonable Function Framework For Device Binding Of Machine Learning Models On Commodity Edge Hardware.

Inventors include Ms. N. Uma; Ms. S. Kavishree; Dr. A. Indumathi; Ms. R. Saktheeswari; Ms. M. Sugacini; Mr. V. Ranjith; Mr. S. Chandrasekar; Mr. P. V. Dhanesh Sivadeep; and Mr. S. Gokhula Anand.

The application for the patent was published on June 19, 2026, under issue no. 25/2026.

Abstract: SoftPUF is a software-based device-binding framework for protecting serialized machine-learning models on commodity edge computing hardware. The framework generates a reproducible device-specific fingerprint from heterogeneous static hardware-associated identifiers and dynamically measured behavioral entropy using a noise-tolerant preprocessing and reconstruction pipeline. The preprocessing pipeline performs normalization, quantization, and canonical encoding of collected entropy measurements prior to error-tolerant reconstruction and cryptographic fingerprint generation. A cryptographic key derived from the reconstructed device-specific fingerprint is used to enforce device-bound access control for encrypted machine-learning models such that successful model decryption, initialization, loading, and inference execution occur only on an authorized physical computing device. During runtime, the fingerprint reconstruction pipeline is re-executed to regenerate the cryptographic key for authenticated in-memory model loading and inference execution without writing decrypted model data to persistent storage. The framework further supports volatile-memory protection mechanisms including memory locking and plaintext memory erasure following runtime initialization. The invention enables stable software-based device binding and runtime cryptographic enforcement for edge-AI deployments without requiring dedicated Physical Unclonable Function circuitry, Trusted Execution Environments, or specialized hardware security modules. Accompanied Drawing [FIGS. 1-7]

Disclaimer: Curated by HT Syndication.