MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641110083 A) filed by Vellore Institute Of Technology on September 14, 2026, for Influence-Trajectory-Based Poisoning Detection With Multi Estimator Stability And Lineage Fingerprinting.
Inventors include Meenakshi; Jaishree Jaikrishnan; Mrinali Prafull Charhate; and Rayban Pranav Mahesh.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: ABSTRACT INFLUENCE-TRAJECTORY-BASED POISONING DETECTION WITH MULTI-ESTIMATOR STABILITY AND LINEAGE FINGERPRINTING The present disclosure provides a method for detecting poisoned training samples in a machine learning model (100). The method comprises storing, by a processor (102), a plurality of checkpoint weight states (110) of the machine learning model (100) captured at successive training epochs. The method comprises computing, for each training sample (112) in a training dataset (114), an influence trajectory (120) comprising a time-ordered sequence of influence scores, wherein each influence score corresponds to a respective checkpoint weight state (110) and represents a gradient alignment between the training sample (112) and a validation dataset (116) at the respective checkpoint weight state (110). The method comprises computing, for each training sample (112), a stability variance score (130) based on variance across a plurality of structurally distinct influence estimators (132), wherein each influence estimator (132) independently estimates gradient influence for the training sample (112). The method comprises identifying one or more poisoned training samples from the training dataset (114) based on the influence trajectory (120) and the stability variance score (130).
Disclaimer: Curated by HT Syndication.
