Federated Learning (FL) is a distributed learning framework that enables collaborative training of a global model across multiple IoT devices while preserving the privacy of local data. However, this collaborative paradigm also introduces security challenges to the global model. Compromised IoT devices can carry out data poisoning attacks on their local data, uploading malicious model gradients to the cloud server and ultimately degrading the global model performance. Most existing data poisoning methods focus primarily on maximizing attack effectiveness, causing a significant drop in global model accuracy, while often neglecting the stealthiness of the attack. This leads to malicious updates being easily detected and filtered, reducing their threat in IoT deployments. To fill this gap, we propose a stealthy data poisoning method based on a dynamic generator. The approach dynamically generates perturbations with gradient-level stealthiness, achieving a trade-off between poisoning effectiveness and stealthiness in realistic IoT settings. We conducted comprehensive experiments on 3 public IoT datasets: with 30% malicious clients, our attack reduces global accuracy by 10.78% on DeepHealth, 6.07% on ChestMNIST and 7.81% on SVHN, achieving degradation comparable to baselines while preserving stealthiness at the gradient level. To counter this threat, We propose PCATrace, a trajectory-based defense that consistently detects and removes malicious clients across different attack ratios, effectively preventing poisoned updates from degrading the global model.

The Industrial Internet of Things (IIoT) increasingly relies on federated learning (FL) to enable collaborative model training without directly sharing raw traffic data across industrial sites. However, in practical IIoT deployments, clients may later request the removal of their data contributions from a trained federated model due to regulatory requirements, such as the General Data Protection Regulation (GDPR), ownership transfer, or internal data-governance policies. Such practical requirements create a strong demand for federated unlearning in IIoT applications. Furthermore, IIoT deployments often exhibit highly imbalanced client data distributions, resulting in substantially different contributions of individual clients to the global model. Nevertheless, most existing federated unlearning methods adopt a uniform unlearning strategy and fail to account for such client-level contribution gaps. To address this issue, we propose DynamicFU, a contribution-aware dynamic federated unlearning framework for IIoT deployments. The proposed method evaluates the target client from parameter-level, data-level, and performance-level perspectives and adaptively determines the unlearning strength by dynamically adjusting the number of unlearning rounds. Experimental results on public IIoT datasets show that DynamicFU substantially improves unlearning efficiency, achieving up to 22.89× speedup over Full Retrain while maintaining comparable effectiveness.