Federated Learning (FL), a secure and emerging distributed learning paradigm, has garnered significant interest in the Internet of Things (IoT) domain. However, it remains vulnerable to adversaries who may compromise privacy and integrity. Previous studies on privacy-preserving FL (PPFL) have demonstrated limitations in client model personalization and resistance to poisoning attacks, including Byzantine and backdoor attacks. In response, we propose a novel PPFL framework, FedRectify, that employs a personalized dual-layer approach through the deployment of Trusted Execution Environments and an interactive training strategy. This strategy facilitates the learning of personalized client features via private and shared layers. Furthermore, to improve model’s robustness to poisoning attacks, we introduce a novel aggregation method that employs clustering to filter out outlier model parameters and robust regression to assess the confidence of cluster members, thereby rectifying poisoned parameters. We theoretically prove the convergence of FedRectify and empirically validate its performance through extensive experiments. The results demonstrate that FedRectify converges 1.47-2.63 times faster than state-of-the-art methods when countering Byzantine attacks. Moreover, it can rapidly reduce the attack success rate to a low level between 10% and 40% in subsequent rounds when confronting bursty backdoor attacks.

Training deep neural networks for classification often includes minimizing the training loss beyond the zero training error point. In this phase of training, a "neural collapse" behavior has been observed: the variability of features (outputs of the penultimate layer) of within-class samples decreases and the mean features of different classes approach a certain tight frame structure. Recent works analyze this behavior via idealized unconstrained features models where all the minimizers exhibit exact collapse. However, with practical networks and datasets, the features typically do not reach exact collapse, e.g., because deep layers cannot arbitrarily modify intermediate features that are far from being collapsed. In this paper, we propose a richer model that can capture this phenomenon by forcing the features to stay in the vicinity of a predefined features matrix (e.g., intermediate features). We explore the model in the small vicinity case via perturbation analysis and establish results that cannot be obtained by the previously studied models. For example, we prove reduction in the within-class variability of the optimized features compared to the predefined input features (via analyzing gradient flow on the "central-path" with minimal assumptions), analyze the minimizers in the near-collapse regime, and provide insights on the effect of regularization hyperparameters on the closeness to collapse. We support our theory with experiments in practical deep learning settings.

Reliable probability estimation is of crucial importance in many real-world applications where there is inherent (aleatoric) uncertainty. Probability-estimation models are trained on observed outcomes (e.g. whether it has rained or not, or whether a patient has died or not), because the ground-truth probabilities of the events of interest are typically unknown. The problem is therefore analogous to binary classification, with the difference that the objective is to estimate probabilities rather than predicting the specific outcome. This work investigates probability estimation from high-dimensional data using deep neural networks. There exist several methods to improve the probabilities generated by these models but they mostly focus on model (epistemic) uncertainty. For problems with inherent uncertainty, it is challenging to evaluate performance without access to ground-truth probabilities. To address this, we build a synthetic dataset to study and compare different computable metrics. We evaluate existing methods on the synthetic data as well as on three real-world probability estimation tasks, all of which involve inherent uncertainty: precipitation forecasting from radar images, predicting cancer patient survival from histopathology images, and predicting car crashes from dashcam videos. We also give a theoretical analysis of a model for high-dimensional probability estimation which reproduces several of the phenomena evinced in our experiments. Finally, we propose a new method for probability estimation using neural networks, which modifies the training process to promote output probabilities that are consistent with empirical probabilities computed from the data. The method outperforms existing approaches on most metrics on the simulated as well as real-world data.

The flow of data across nodes has become the dominant feature of data sharing in distributed environments with increasingly blurred boundaries, where it is crucial to maintain data access dynamic, trusted, and efficient. However, traditional centralized access control models are not only difficult to apply in distributed environments but also ignore trusted verification of authorized entities. What’s worse, existing access control models rarely consider themselves security, lack independence, at a high risk of being bypassed or tampered with. Thus, we propose in this paper a distributed, dynamic, and trusted access control model, DDTAC-BSS, where the standard Attribute-Based Access Control (ABAC) architecture is modified and extended. To reduce the attack surface, we separate policy enforcement point (PEP) from other core components, they are located in the node system and access control system, respectively. Then, the access control entry point (ACEP) is added as the only interface for the node system to interact with the access control system. Subsequently, the model introduces the entity trusted assessment mechanism to improve the trustworthiness of access control services. Driven by the dynamic attributes, our model can achieve dynamic trusted authorization and fine-grained access control. Moreover, we implement a lightweight, independent, and distributed security subsystem to achieve unified management of policies and decision-making autonomy by message-driven. By considering the independence of the security subsystem, a trusted operating environment is built based on Trusted Execution Environment (TEE) to ensure the security of the access control mechanism itself. The security of our model is proved rigorously based on the non-interference theory. Comprehensive experiments and comparisons have demonstrated the superior functionality, comparable performance, and strong security of our model.