Federal learning backdoor attack defense method based on DAGMM

Abstract

The invention discloses a federal learning backdoor attack defense method based on DAGMM, and the method comprises the following steps: (1) a client receives a global model, and trains and uploads a local model and a corresponding neuron activation condition; (2) the server receives the update and calculates the loss of the corresponding client by using the DAGMM; and (3) defense is performed based on multi-round reconstruction errors. According to the invention, the model can be effectively protected from backdoor attack.

Description

technical field [0001] The invention relates to the technical field of backdoor attack defense, in particular to a DAGMM-based federated learning backdoor attack defense method. Background technique [0002] Federated learning has been proposed to facilitate joint model training utilizing data from multiple clients, where the training process is coordinated by a central server. During the whole process, the client's data is kept locally, and only the model parameters are communicated between clients through the parameter server. [0003] A typical training iteration works as follows: First, the central server sends each client the latest global model. Each client then locally updates the model with local data and reports the updated model to the parameter server. Finally, the server performs model aggregation on all submitted local updates to form a new global model that performs better than models trained with data from any single client. [0004] Compared to the alterna...

AI Technical Summary

Problems solved by technology

However, this way of pruning relies on a reliable source of "clean" data, which is not guaranteed in federated learning scenarios

Images