Abstract
Adversarial attacks in network systems manipulate data to deceive networks, impacting domains like image recognition and natural language processing. Attackers exploit algorithm weaknesses, creating misleading inputs that lead to inaccurate outcomes. Defending against these evolving threats is demanding, but federated learning (FL) enhances model generalization by learning from diverse data sources. FL also decreases single points of failure, complicating adversaries’ efforts to manipulate the global model effectively. Here, a new model named dollmaker weighted moving average–based quantum recurrent neural network is developed for adversarial attack detection. Here, FL is composed of nodes and servers. At every node, the local training is conducted based on local data. Thereafter, model aggregation is performed at the server, sending the data back to the local nodes and the same process is iterated at every epoch. In the training model, the input data acquired are preprocessed using the min–max normalization technique. Then, the adversarial attack detection is conducted by a quantum recurrent neural network, which is trained utilizing dollmaker weighted moving average. Here, aggregation at the server and local update are modified utilizing the average method. It is identified that the dollmaker weighted moving average–based quantum recurrent neural network has achieved a normalized mean square error (MSE) of 0.081, an accuracy of 90.90%, and a loss function of 0.091. When considering accuracy metrics, the performance improvement achieved by the proposed method is 11.914%, 8.339%, 8.306%, 8.119%, 5.490%, and 2.739% higher than the existing methods. These results indicate that the dollmaker weighted moving average–based quantum recurrent neural network provides robust and reliable detection of adversarial attacks in distributed network systems.