MUMBAI, India, Aug. 17 -- Intellectual Property India has published a patent application (202641094909 A) filed by Pes University on August 05, 2026, for Method And System For Defending Deep Learning Models Against Backdoor Attacks.
Inventors include Nagasundari S; and Prasad B Honnavalli.
The application for the patent was published on August 14, 2026, under issue no. 33/2026.
Abstract: The present disclosure relates to a system (100) and method (600) for defending a deep learning model (132) against backdoor attacks. The method (600) may include training (602) the deep learning model on a dataset including clean and poisoned samples, computing (604) suspicion scores (210) for each sample by jointly evaluating training loss (204), gradient norm (206), and prediction confidence (208) using a multi-signal behavioral analysis unit (112). A dynamic isolation unit (114) may determine an adaptive isolation threshold (212) and isolation ratio (214) to segregate samples into a clean and suspected poisoned subset (218). An iterative re-isolation unit (116) may periodically re-evaluate the clean subset to capture undetected poisoned samples (304). A hybrid adversarial unlearning unit (118) may perform a hybrid unlearning operation (408), and finally the deep learning model is stabilized (614), thereby generating a backdoor-resilient deep learning model (412).
Disclaimer: Curated by HT Syndication.
