MUMBAI, India, Sept. 14 -- Intellectual Property India has published a patent application (202641105133 A) filed by V. Kalpana; Chandrakala V; Dr. B. Gunasundari; Dr. Sarojini Yarramsetti; Dr. Yuvaraj. B; and S. Kumaresan on September 01, 2026, for Ai-Based Runtime Model Integrity Verification Using Behavioral And Activation Fingerprints.

Inventors include V. Kalpana; Chandrakala V; Dr. B. Gunasundari; Dr. Sarojini Yarramsetti; Dr. Yuvaraj. B; and S. Kumaresan.

The application for the patent was published on September 11, 2026, under issue no. 37/2026.

Abstract: AI-Based Runtime Model Integrity Verification Using Behavioral and Activation Fingerprints proposes a lightweight framework for detecting unauthorized modification, tampering, or replacement of deployed deep learning models during runtime. The proposed system uses activation fingerprints and behavioral fingerprints to continuously verify whether the executing model remains consistent with its trusted reference state. During a secure enrollment phase, a pretrained model is profiled using a representative validation dataset and intermediate-layer activation statistics are extracted using Hook-based Activation Monitoring. These features are compressed using Principal Component Analysis (PCA) and stored as a cryptographically protected reference fingerprint using SHA-256. During inference the framework captures runtime activations and output behavior and computes deviation shores using Mahalanobis Distance and Cosine Similarity. An Isolation Forest anomaly-detection algorithm combines these deviations to identify abnormal model behavior. A Weighted Fingerprint Integrity Score (WFIS) is then calculated from activation and behavioral discrepancies, and an adaptive threshold determines whether the model is Trusted, Suspicious, or Compromised. For stronger protection, the reference fingerprints and verification logs can be secured using AES-256 encryption and blockchain-based immutable logging. The proposed approach provides continuous, model-aware runtime integrity verification without requiring access to the original model source code or retraining the model, making it suitable for edge Al, healthcare AI, autonomous systems, and security-critical AT deployments.

Disclaimer: Curated by HT Syndication.