Abstract: The system comprises a FL-based generative adversarial network (GAN) for generating adversarial examples, wherein the GAN includes a generator for generating the adversarial examples and a discriminator for distinguishing the adversarial examples from the original data, wherein the FL network includes multiple clients, each having a local dataset and a local DNN model, and a central server for coordinating the training process; a DNN for classifying data, where the DNN is trained using the generated adversarial examples, wherein the training process includes exchanging the model updates between the client’s server and the central server; an evaluation module for measuring the adversarial accuracy and adversarial robustness of the DNN using appropriate metrics, including the adversarial accuracy, the adversarial loss, and the robustness to perturbations; and an adjustment module for adjusting the architecture or parameters of the DNN based on the evaluation results to improve its adversarial robustness.