Scholarlink

Federated Learning (FL) offers a privacy-preserving paradigm for decentralized machine learning, yet its performance is fundamentally challenged by the Non-Independent and Identically Distributed (Non-IID) nature of client data. This paper introduces AFedGAN, a novel, client-centric framework that mitigates data heterogeneity through adaptive generative augmentation. Each client pre-trains a local conditional GAN (cGAN) and, during each FL round, identifies its worst-performing class based on local model feedback. It then generates targeted synthetic samples for this class. A crucial confidence-based filtering mechanism, guided by the client’s current classifier, ensures that only high-quality synthetic data is retained for local training. This entire process is performed locally, preserving data privacy without reliance on public datasets or server-side coordination. We conduct extensive experiments on pathologically Non-IID splits of EMNIST and CIFAR-10. The results demonstrate that AFedGAN significantly outperforms key baselines including FedAvg, FedProx, traditional augmentation, and non-adaptive FedGAN across accuracy, fairness, and robustness. For instance, on EMNIST, AFedGAN achieves a test accuracy of 0.8891, surpassing FedAvg (0.8329) and FedProx (0.8451). On the more challenging CIFAR-10 dataset, AFedGAN also achieves the highest accuracy of 0.7551, validating its effectiveness and generalizability.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AFedGAN: Adaptive Federated Learning with Generative Adversarial Networks for Non-IID Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xuyang Zhang,
  • Hua Jin,
  • Peiyuan Guo

摘要

Federated Learning (FL) offers a privacy-preserving paradigm for decentralized machine learning, yet its performance is fundamentally challenged by the Non-Independent and Identically Distributed (Non-IID) nature of client data. This paper introduces AFedGAN, a novel, client-centric framework that mitigates data heterogeneity through adaptive generative augmentation. Each client pre-trains a local conditional GAN (cGAN) and, during each FL round, identifies its worst-performing class based on local model feedback. It then generates targeted synthetic samples for this class. A crucial confidence-based filtering mechanism, guided by the client’s current classifier, ensures that only high-quality synthetic data is retained for local training. This entire process is performed locally, preserving data privacy without reliance on public datasets or server-side coordination. We conduct extensive experiments on pathologically Non-IID splits of EMNIST and CIFAR-10. The results demonstrate that AFedGAN significantly outperforms key baselines including FedAvg, FedProx, traditional augmentation, and non-adaptive FedGAN across accuracy, fairness, and robustness. For instance, on EMNIST, AFedGAN achieves a test accuracy of 0.8891, surpassing FedAvg (0.8329) and FedProx (0.8451). On the more challenging CIFAR-10 dataset, AFedGAN also achieves the highest accuracy of 0.7551, validating its effectiveness and generalizability.