FedDyna: Privacy-Preserving Federated Learning with Dynamic Noise Adaptation and Structural Bias Alignment for Non-IID Environments

In Internet of Things (IoT) scenarios, federated learning (FL) facilitates data value mining through distributed collaborative learning while preserving user privacy. However, the performance of existing privacy-preserving methods in federated learning systems is greatly reduced due to the problem of non-independent and identically distributed (non-IID) data. In this paper, we study the privacy preservation problem of FL in non-IID data environments and propose FedDyna, a novel framework that uses an adaptive noise injection mechanism to enhance privacy. A theoretical analysis reveals an upper bound on the convergence of additive noise under non-IID and proves the limitations of the traditional static noise strategy. FedDyna uses a decaying noise injection method that dynamically correlates with model biases to synchronize structured privacy preservation of the local parameter space with global consistency maintenance in the client training phase. Meanwhile, through a locally drifting tracking mechanism constrained by matrix trace, we explicitly decouple model parameter deviations and perform dynamic correction, combined with gradient variance regularization to suppress divergence in local updates. Finally, through simulation experiments conducted on two benchmark datasets, MNIST and CIFAR-10, we effectively demonstrate the significant divergence between dynamic and static noise under non-IID settings.