MUMBAI, India, Aug. 17 -- Intellectual Property India has published a patent application (202641093915 A) filed by Vardhaman College Of Engineering on August 03, 2026, for Adversarial Machine Learning Defense Framework Using Meta-Learning For Attack Adaptation.
Inventors include Mr. N S S S Girish Kumar; Dr. R Karthikeyan; Dr. Ramachandro Majji; Mr. Ramachandra Rao Moka; Mr. Yogash Chandra Joshi; and Mr. Abhishek Dixit.
The application for the patent was published on August 14, 2026, under issue no. 33/2026.
Abstract: Adversarial Machine Learning Defense Framework Using Meta-Learning for Attack Adaptation is the proposed invention. The present invention provides an adversarial machine learning defence framework based on Model-Agnostic Meta-Learning (MAML) to adaptively mitigate the evolving adversarial attacks. The framework consists of an input acquisition module, feature extraction unit, attack characterisation engine, MAML-based meta-learning module, adaptive defence optimiser, robust classifier, confidence estimation module, and continual knowledge repository. The MAML algorithm is trained on meta-training with generalised initialisation parameters for different attack scenarios, and it can adapt to unseen attack scenarios with limited gradient updates. We leverage a framework that performs real-time dynamic optimisation of the defence parameters upon the discovery of adversarial perturbations without retraining the full model, thus reducing the computational complexity and improving the real-time performance. The continual learning repository incrementally integrates new attack knowledge and preserves the previously learned defensive capabilities. The invention proposed herein provides significantly higher robustness against white-box, black-box, transfer and zero-day attacks, while maintaining prediction accuracy, computational efficiency, scalability and reliability for deployment in safety-critical artificial intelligence applications.
Disclaimer: Curated by HT Syndication.
