Federated learning (FL) has become a promising framework for collaborative model training on devices while preserving privacy. However, despite its significant potential, it faces notable cyber threats, such as poisoning attacks and codenamed Byzantine clients. These threats have the potential to significantly degrade the global model by jeopardizing the integrity of the collaborative model training process. Whilst previous research has addressed the detection and elimination of malicious gradients from Byzantine clients, it has also shown that model poisoning attacks can evade most statistical defence approaches relying on metrics such as median and distance. To address the challenge posed by poisoning attacks, we introduce a novel approach called Layerwise Gradient Purify (LGP), which aims to remove any harmful gradients before the global aggregation process. It is comprised of two closely-related stages. The first stage focuses on pruning gradients at the layer level using the Median Absolute Deviation (MAD) pruning criterion. In the second stage, statistical features are extracted from the pruned gradients layer-by-layer and then clustered into honest and malicious categories. The proposed methodology treats each layer of the model across all clients as a probability distribution, employing hierarchical clustering to differentiate between malicious and honest clusters. Moreover, we introduce a new innocent criterion for selecting honest clusters, relying on reputation scores and gradient deviations from the global model. Extensive experiments were conducted employing diverse deep learning models, including CNN, RNN, and MLP, across a spectrum of datasets, including Cifar-10, AG-News, MNIST and ToN-IoT. The experiments evaluated the resilience of state-of-the-art approaches against recently introduced attacks. The numerical results demonstrate that theLGPapproach is effective and superior.

Sixth-generation (6G) wireless networks are set to transform the Internet of Things (IoT) by enabling faster, smarter, and more connected systems. These networks will bring together a wide range of devices, including cars, robots, industrial machines, and smartphones, to support edge intelligence and real-time decision-making. Federated learning (FL) supports this shift by allowing devices to collaboratively train models without sharing raw data, which helps to protect user privacy. Despite its advantages, FL faces significant security challenges, including poisoning attacks and Byzantine clients, both of which can compromise the training process and degrade the accuracy and reliability of the global model. Although existing methods can detect malicious updates, many advanced attacks still bypass statistical defenses relying on metrics such as median and distance. Thus, developing an FL system that ensures both reliable decision-making and privacy and security guarantees in IoT networks remains a significant challenge. This study introduces a Digital Twin-driven Blockchain-enabled Federated Learning (DT-BFL) framework designed for IoT networks. The framework creates a digital representation of the IoT environment to support secure and decentralized edge intelligence using blockchain and federated learning technologies. DT-BFL is built to detect and filter out potentially poisoned model updates from malicious participants. This is achieved through a new smart contract-enabled decentralized aggregation method called Local Updates Purify (LUP). LUP uses a two-stage filtering process: First, it applies Median Absolute Deviation (MAD) to initially remove outliers, then uses statistical features and clustering to separate honest from malicious updates before aggregating the global model. It also assigns a Trust Score (TS) to each participant based on how much their updates differ from the global model and then uses a genuine criterion to select honest clients by evaluating trust scores, update similarity, and deviation from the global model. Experimental results show that DT-BFL effectively defends against various poisoning attacks on datasets like MNIST, ToN-IoT, and CIFAR-10 using models such as CNN, MLP, ResNet, and DenseNet, and maintains high accuracy even when 50% of the clients are malicious. Using a permissioned blockchain further secures the system by enabling aggregation of the decentralized model and authentication of clients through smart contracts. The source code is available on https://github.com/UNSW-Canberra-2023/LUP .

Federated learning (FL) enables distributed joint training of machine learning (ML) models without the need to share local data. FL is, however, not immune to privacy threats such as model inversion (MI) attacks. The conventional FL paradigm often uses privacy-preserving techniques, and this could lead to a considerable loss in the model’s utility and consequently compromised by MI attackers. Seeking to address this limitation, this paper proposes a robust variational encoder-based personalised FL (RVE-PFL) approach that mitigates MI attacks, preserves model utility, and ensures data privacy. RVE-PFL comprises an innovative personalised variational encoder architecture and a trustworthy threat model-integrated FL method to autonomously preserve data privacy, and mitigate MI attacks. The proposed architecture seamlessly trains heterogeneous data at every client, while the proposed approach aggregates data at the server side and effectively discriminates against adversarial settings (i.e., MI); thus, achieving robustness and trustworthiness in real-time. RVE-PFL is evaluated on three benchmark datasets, namely: MNIST, Fashion-MNIST, and Cifar-10. The experimental results revealed that RVE-PFL achieves high accuracy level while preserving data and tuning adversarial settings. It outperforms Noising before Model Aggregation FL (NbAFL) with significant accuracy improvements of 8%, 20%, and 59% on MNIST, Fashion-MNIST, and Cifar-10, respectively. These findings reinforce the effectiveness of RVE-PFL in protecting against MI attacks while maintaining the model’s utility. The source code for RVE-PFL can be found on GitHub 1.