Federated learning enables privacy-preserving model training but remains vulnerable to data poisoning, where malicious clients manipulate local data to induce targeted misclassification. Existing defenses often assume that (i) benign updates are mutually consistent and (ii) labeled data are available for anomaly detection. These assumptions rarely hold under realistic non-IID distributions and unlabeled settings, leading to false detection and degraded robustness. To address these challenges, we propose D3FL, a label-free and heterogeneity-robust defense framework that detects anomalies through client–reference discrepancy analysis. Instead of inter-client comparisons, D3FL distills a student model from the global model and consensus client predictions, constructing a trusted reference model that encapsulates reliable client knowledge. A multi-perspective inconsistency evaluation then quantifies prediction anomalies, feature deviations, and parameter discrepancies, accurately distinguishing benign heterogeneity from malicious manipulation. To remove label dependence, a voting-based consensus mechanism selects high-confidence client predictions to refine the reference model via distillation. Moreover, a dynamically adaptive aggregation strategy integrates real-time consistency with historical trust, continuously adjusting client contributions to maintain robustness throughout training. Extensive experiments on Fashion-MNIST and CIFAR-10 under non-IID settings demonstrate that D3FL achieves strong robustness against data poisoning without requiring any labeled clean data.

ABSTRACT The continuous collection of streaming data in the Internet of Things (IoT) may compromise user privacy, as such data often originates from personal information. Local differential privacy (LDP) is a novel privacy notion that offers a strong privacy guarantee to all users without relying on a trusted data collector. However, existing LDP‐based studies mainly focus on static scenarios or perturbation of data points at a single timestamp without sufficiently considering data sparsity, which adds excessive noise and leads to low utility. Therefore, we propose a Framework for Sparse Streaming Data Collection based on Local Differential Privacy (SS‐LDP), which aims to provide high utility at each timestamp while satisfying ‐event LDP. One component is the introduction of an upper‐bound optimization mechanism, which reduces the noise scale by combining error minimization with the gradient descent method. Another component of SS‐LDP targets the efficient management of privacy resources through two specific strategies. First, significant changes in streaming data are captured by calculating differences between the latest few data points, thereby conserving the privacy budget. Second, an improved sparse privacy budget allocation mechanism quantifies data sparsity at each timestamp using the moving average method, enabling efficient allocation of the privacy budget for each timestamp. SS‐LDP is evaluated using two real‐world datasets and compared with four baseline methods that satisfy ‐event privacy. Extensive experiments and theoretical analyses are conducted to demonstrate the superiority of our framework.

Federated learning (FL), a cutting-edge distributed machine learning training paradigm, aims to generate a global model by collaborating on the training of client models without revealing local private data. The cooccurrence of non-independent and identically distributed (non-IID) and long-tailed distribution in FL is one challenge that substantially degrades aggregate performance. In this paper, we present a corresponding solution called federated dual-decoupling via model and logit calibration (FedDDC) for non-IID and long-tailed distributions. The model is characterized by three aspects. First, we decouple the global model into the feature extractor and the classifier to fine-tune the components affected by the joint problem. For the biased feature extractor, we propose a client confidence re-weighting scheme to assist calibration, which assigns optimal weights to each client. For the biased classifier, we apply the classifier re-balancing method for fine-tuning. Then, we calibrate and integrate the client confidence re-weighted logits with the re-balanced logits to obtain the unbiased logits. Finally, we use decoupled knowledge distillation for the first time in the joint problem to enhance the accuracy of the global model by extracting the knowledge of the unbiased model. Numerous experiments demonstrate that on non-IID and long-tailed data in FL, our approach outperforms state-of-the-art methods.