MUMBAI, India, Sept. 7 -- Intellectual Property India has published a patent application (202541058993 A) filed by Sisa Information Security Pvt Ltd on June 19, 2025, for Ai-Based System And Method For Hardening Response Generation Process Of Ai-Models Via Targeted Layer Fine-Tuning.
Inventors include Anjan Krishnamurthy; Renju Varghese; Mahesh R; and Bharath M.
The application for the patent was published on September 04, 2026, under issue no. 36/2026.
Abstract: ABSTRACT AI-BASED SYSTEM AND METHOD FOR HARDENING RESPONSE GENERATION PROCESS OF AI-MODELS VIA TARGETED LAYER FINE- TUNING The present invention discloses an AI-based system (102) and a method (500) for hardening response generation process of AI models. The process begins by applying (504) adversarial prompts into the AI models to generate first responses. The process followed by analyzing (510) transformer layers of AI models, which provide contradict responses based on set of broken vulnerable prompt-response pairs by systematic recursive layer pruning analysis. The process further includes pruning (512) a transformer layer to adapt the AI models to generate second responses corresponding to the adversarial prompts. The process further includes assigning (516) scores to the second responses based on safety level of second responses by ML classifier upon pruning of the transformer layers. The process further includes isolating (518) transformer layers that are associated with contradict responses. The process further includes fine-tuning (520) the isolated transformer layers of the AI models to generate hardened responses. FIG. 5
Disclaimer: Curated by HT Syndication.
