Backdoor behavior detection method and system based on weight tensor

CN122247768APending Publication Date: 2026-06-19STATE GRID ZHEJIANG ELECTRIC POWER CO LTD +1

Patent Information

Authority / Receiving Office: CN · China
Patent Type: Applications(China)
Current Assignee / Owner: STATE GRID ZHEJIANG ELECTRIC POWER CO LTD
Filing Date: 2026-05-22
Publication Date: 2026-06-19

AI Technical Summary

Technical Problem

Existing deep neural network backdoor detection methods are difficult to meet the requirements of not relying on training data and trigger patterns in power grid scenarios. They are also computationally expensive, lack accuracy and robustness, and cannot quickly screen for diverse attacks.

Method used

By obtaining weight tensors from deep neural network models, standardizing them, and then extracting features from dimensions such as independence, correlation, and common structure, the final classifier is used to determine whether the model has backdoor behavior. Independent vector analysis, multi-set canonical correlation analysis, and parallel factor analysis algorithms are used for feature extraction, and interpretable AI technology is introduced for attribution analysis.

Benefits of technology

It achieves efficient and accurate detection of backdoor behavior in deep neural networks without relying on training data and trigger patterns, improving detection accuracy and robustness, providing an interpretable chain of evidence for decision-making, and is suitable for power grid security scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

Figure CN122247768A_ABST

Patent Text Reader

Abstract

This invention discloses a backdoor behavior detection method and system based on weight tensors. The method includes obtaining target weights corresponding to several network layers from a deep neural network model to be detected; standardizing the target weights to obtain corresponding weight tensors; extracting features from the weight tensors from different dimensions to obtain feature vectors of the weight tensors; combining the feature vectors to obtain a feature matrix of the weight tensors; and using a final classifier to classify the feature matrix to determine whether the deep neural network model to be detected exhibits backdoor behavior. This method can capture abnormal patterns in model weights from different dimensions without relying on training data and trigger patterns, thus more comprehensively and efficiently capturing weight perturbations caused by different backdoor implantation techniques, improving the accuracy and robustness of deep neural network backdoor detection.

Need to check novelty before this filing date? Find Prior Art