MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202611083599 A) filed by Delhi Technological University on July 07, 2026, for System And Method For Explainable Ai-Guided Semantic Watermarking Of Deep Neural Networks Dnn.
Inventors include Gupta, Nikhil; Jain, Agamya; Ankur; and Kumar, Rajeev.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: A system (102) and a method (300) for embedding an ownership watermark into a DNNmodel is disclosed. The system (102) comprises a controller (104) comprising a processor (110) and a memory (106) configured to obtain a trained reference classifier, determine a confusion matrix representing misclassification relationships among classes, and generate class-specific target sets and conditioning vectors. The processor further generates attribution information using an Explainable Artificial Intelligence (XAI) technique, determines sensitivity scores to select training samples for watermark embedding, and generates spatial masks identifying important image regions. A trigger generator produces conditioning vector-based perturbations, which are masked and embedded into chrominance frequency subbands of selected training samples to generate watermarked training samples. A target DNNmodel is trained using the watermarked and non-watermarked training samples such that verification inputs induce predictions corresponding to classes within the target sets.
Disclaimer: Curated by HT Syndication.
