MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202611109968 A) filed by Vikas Somani; Vikas Joshi; Pallavi Krishna Purohit; Richa Sharma; Snehlata Ajmera; Sweety Singhal; and Vartika Vyas on September 14, 2026, for Privacy-Preserving Deep Neural Network Intrusion Detection System Using Differential Privacy And Evolutionary Optimization.
Inventors include Vikas Somani; Vikas Joshi; Pallavi Krishna Purohit; Richa Sharma; Snehlata Ajmera; Sweety Singhal; and Vartika Vyas.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: A computer-implemented privacy-preserving intrusion detection system and method are disclosed for detecting malicious activity happening within a network while minimizing leakage of privacy information within the machine-learning training data. Information about the network traffic is pre-processed and fed to one or more deep neural-network architecture (DNN) models trained to classify intrusions. An optimization level differential-privacy mechanism or a teacher-ensemble privacy mechanism is added to the model generation to ensure privacy protection. For optimization level, the contribution of each training sample is clamped and there is a controlled random jitter during the updating of neural network parameters. Cumulative privacy expenditure is determined using a privacy-accounting mechanism. In an ensemble setting, private training set is partitioned into mutually disjointed sets and independent teacher networks are trained with their respective sets, teacher predictions are privacy-preservingly aggregated and a student intrusion detection model is trained even without access to private training set of teachers with privatized labels. The teacher networks can be homogenous or heterogeneous. An evolutionary optimization engine is used to search model-training and privacy parameters given a plurality of model configurations. The fitness of the candidate is based on the amount of utility intruded and the amount of privacy expenditure, thus allowing choosing a configuration that meets a specific privacy-utility equation. Evolutionary learning modifies learning parameters, architecture parameters, gradient-limiting parameters, noise parameters, batch, regularization parameters or a combination. The chosen intrusion-detection model is put in place to check later network traffic, and provide indication of either benign or malicious activity. The disclosed architecture thus provides measurable privacy protection as part of the technical selection and training of a network-security classifier, instead of as a post-training procedure.
Disclaimer: Curated by HT Syndication.
