Heterogeneous federated learning backdoor attack detection method and device, equipment and medium

By calculating parameter offsets and dynamic adjudication thresholds in federated learning to identify potential malicious actors, removing their parameters, and weighting and aggregating benign client parameters, the problem of model update differences caused by non-independent and identically distributed client data and the difficulty in distinguishing them from malicious attacks is solved, thereby improving the security and generalization performance of the global model.

CN122372323APending Publication Date: 2026-07-10PURPLE MOUNTAIN LAB
0 Cites 0 Cited by

Patent Information

Application Number
CN202610739169.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-26
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

In the distributed training process of federated learning, the non-independent and identically distributed nature of client data makes it difficult to distinguish between model update differences and malicious attack behavior. Existing defense mechanisms have failed to effectively deal with this, resulting in a decrease in global model convergence accuracy and insufficient security.

Method used

By calculating the parameter offset of the client model parameters relative to the previous round of the global model, the adjudication threshold is dynamically determined, potential malicious actors are identified and their model parameters are removed. Differential weights are used to aggregate benign client parameters to generate global model parameters for the new training round.

Benefits of technology

It achieves accurate identification and defense against covert backdoor attacks and data poisoning while ensuring the convergence accuracy of the global model, thereby improving the security and robustness of the federated learning system and optimizing accuracy and generalization performance.

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

Abstract

The invention relates to the technical field of federated learning, and discloses a heterogeneous federated learning backdoor attack detection method, device and equipment and a medium, which are applied to a central server, and the method comprises the following steps: issuing global model parameters to a plurality of clients; model parameters uploaded by the clients are received, and the parameter offset of the model parameters of the clients relative to the global model parameters of the last training round is determined; determining a dynamic judgment threshold based on the parameter offsets of all the clients, and judging the clients with the parameter offsets exceeding the dynamic judgment threshold as potential malicious executors; and when it is detected that the potential malicious executors are abnormal behaviors, the potential malicious executors are eliminated, and model aggregation is performed on the residual model parameters uploaded by the benign client to generate global model parameters of a new training round. According to the technical scheme provided by the invention, on the premise that the convergence precision of the global model is guaranteed, accurate recognition and defense of hidden backdoor attacks or data poisoning behaviors can be realized.
Need to check novelty before this filing date? Find Prior Art