Time dimension segmentation backdoor attack method for federal neuromorphic learning system

By segmenting neuromorphic data into multiple time slices according to the time axis and embedding local triggers, a time-dimensional segmentation backdoor attack method was proposed for the federated neuromorphic learning system, which solved the problem of lack of attack methods for time-dimensional poisoning data sets in FedNL, and achieved efficient backdoor attacks, significantly improving the attack success rate and concealment.

CN119740231AInactive Publication Date: 2025-04-01NO 30 INST OF CHINA ELECTRONIC TECH GRP CORP +1
0 Cites 0 Cited by

Patent Information

Application Number
CN202411772055.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-04
Publication Date
2025-04-01
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing backdoor attack research mainly focuses on deep neural networks (DNNs), and there is no effective technology for the time-dimensional poisoning data set of neuromorphic data in the federated neuromorphic learning system (FedNL), resulting in the lack of time-dimensional backdoor attack methods for FedNL in the federated learning process.

Method used

A backdoor attack method for the time dimension segmentation of federated neuromorphic learning system is proposed. By segmenting the neuromorphic data into multiple time slices according to the time axis and embedding local triggers in each time slice, the malicious client uploads model parameters with local triggers to the central server during training to form a global model. In the inference phase, a time sequence global trigger is formed by combining multiple local triggers, resulting in misclassification of the model when inputting a specific time series.

Benefits of technology

Through the time-dimensional segmentation of backdoor attack method, the concealment and success rate of the attack are enhanced, the robustness and concealment of backdoor attacks are effectively improved, and distributed timing backdoor attacks can be successfully implemented in the federated neuromorphic learning system.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention relates to a time dimension segmentation backdoor attack method for a federal neuromorphic learning system, and the method comprises the steps: time slice segmentation: segmenting neuromorphic data into a plurality of time slices according to a time axis; local trigger injection: embedding a local trigger in a respective data time slice by each malicious client by controlling a plurality of malicious clients; model training and aggregation are carried out, in the training process, a malicious client side adopts a local data training model embedded with a local trigger, and the local data training model and model parameter updating of a normal client side are uploaded to a central server together for parameter aggregation, so that a global model is formed; timing sequence combination attack is carried out in the reasoning stage, in the reasoning stage, an attacker forms a timing sequence global trigger by combining a plurality of local triggers, and when the global model receives input with the timing sequence global trigger, a classification result is misclassified to a target class. Compared with the prior art, the attack concealment can be effectively enhanced, and the attack success rate is remarkably improved.
Need to check novelty before this filing date? Find Prior Art