MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641126826 A) filed by Vellore Institute Of Technology on October 24, 2026, for System And Method For Screening Neural-Network Model Artifacts For Concealed Code Before Deployment.
Inventors include Balaji Gn; Erukulla Yashwanth; Shreeyash Santosh Ugale; and Maddipati Puneet Janakiram.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: ABSTRACT SYSTEM AND METHOD FOR SCREENING NEURAL-NETWORK MODEL ARTIFACTS FOR CONCEALED CODE BEFORE DEPLOYMENT A computer-implemented method (200) screens an untrusted neural-network model artifact before deployment. The model artifact is received (202) and triaged (204) through a weight-access interface exposing tensor shapes, data types, tensor values and configuration metadata, without invoking a forward or generation function of the neural-network model. One or more signals, each producing a score, a confidence and evidence, are computed (206) and combined (208) into a memorization-risk verdict. Files are reconstructed (210) from the model weights as recovered code treated as untrusted input. The recovered code is statically analyzed (212) and executed (214) within an isolated, network-disabled sandbox environment while runtime behavior is recorded. Static and dynamic findings are fused (216) into a benign, suspicious or malicious risk verdict having a score and a confidence, the risk verdict being reported (218) with the memorization-risk verdict for an allow, quarantine or block decision. (Fig. 2)
Disclaimer: Curated by HT Syndication.
