Method, device and program product for dynamic protection of anti-physical attack edge AI model

By collecting data streams and physical environment parameters in the edge federated learning system, using spiking neural networks and graph neural networks for threat assessment and anomaly detection, and combining neuromorphic physical non-cloning function verification model update packages to dynamically adjust system power consumption, the security and energy efficiency issues of the edge federated learning system in the face of physical attacks are solved, achieving real-time perception and low-power operation.

CN122179140APending Publication Date: 2026-06-09CHINA TOWER CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHINA TOWER CO LTD
Filing Date
2026-02-02
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing edge federated learning systems lack physical layer trusted verification when facing physical attacks and advanced logic attacks, making it difficult to identify covert timing spoofing behaviors. Furthermore, security policies and energy efficiency management are disconnected, making it impossible to support long-term low-power operation.

Method used

By collecting input data streams and physical environment parameters, a spiking neural network is used for threat assessment, a time-series graph structure is constructed for anomaly detection, and a graph neural network is used to identify abnormal behavior. The physical layer verification of the model update package is performed by combining neuromorphic physical non-cloning functions and quantum random numbers, and the system power consumption and response strategy are dynamically adjusted.

Benefits of technology

It enables real-time detection and identification of physical layer attacks, prevents firmware tampering and hardware cloning, ensures security while achieving adaptive energy efficiency optimization, and supports long-term low-power and reliable operation of edge devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122179140A_ABST
    Figure CN122179140A_ABST
Patent Text Reader

Abstract

The present disclosure provides an anti-physical attack edge AI model dynamic protection method, device and program product, belonging to the technical field of network security. The anti-physical attack edge AI model dynamic protection method of the present disclosure collects input data stream and physical environment parameters; threat assessment is performed on the input data stream based on a pulse neural network to generate a dynamic defense strategy; a time sequence diagram structure is constructed, and abnormal behavior detection is performed on the time sequence diagram structure through a graph neural network; an AI model is trained locally, and the obtained model gradient is uploaded after differential privacy disturbance is applied; a model update package is received, and a neural pseudo-state physically unclonable function and quantum random number are used to perform physical layer verification on the model update package; and the attack level determined according to the abnormal behavior detection result or the threat assessment result is used to dynamically adjust the system power consumption and the response strategy.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Decentralized federated learning-based poisoning attack dynamic defense method

    CN120602123A

  • Security defense method for edge federated network attack

    CN120785664A