In next-generation 6G Internet-of-Things (IoT) networks, semantic communication has emerged as a key paradigm that transforms raw data into high-level feature representations, thereby reducing communication overhead while enhancing interpretability. When combined with federated learning (FL), these semantic embeddings enable decentralized model training without centralizing raw data, preserving user privacy, and supporting large-scale collaboration. However, semantic features may inadvertently encode sensitive information or act as adversarial triggers, introducing new privacy risks that current unlearning techniques fail to address. To overcome this challenge, we propose Selective Semantic Feature Unlearning (SSFU), a novel framework that performs unlearning at the feature level rather than at the client level. SSFU employs an ensemble-based risk scoring mechanism to identify high-risk latent components, followed by gradient ascent and semantic masking to remove their influence. Unlike existing methods that depend on costly retraining or full client exclusion, SSFU preserves benign semantic knowledge and allows training to continue with minimal disruption. The framework guarantees bounded convergence, and empirical results on benchmark datasets show that SSFU effectively eliminates sensitive features while maintaining predictive accuracy. SSFU thus represents a robust, privacy-preserving FL framework tailored for semantic communication in 6G IoT systems.

The proliferation of Internet of Things (IoT) devices has intensified the need for scalable and energy-efficient federated learning (FL). While Hierarchical Federated Learning (HFL) improves scalability by adding an edge aggregation tier, it still suffers from high communication costs, slow convergence, and degraded accuracy under non-IID data. Existing methods such as quantization, sparsification, and static pruning alleviate specific bottlenecks but fail to jointly optimize efficiency, robustness, and accuracy. This paper proposes an intent-driven dual-layer model pruning framework for HFL, where an Energy Management System (EMS) and an Intent-driven Pruning Orchestrator (IDPO) dynamically translate system-level intents (e.g., energy minimization or accuracy preservation) into pruning actions at both edge and cloud layers. Experiments on MNIST, CIFAR-10, and FEMNIST show up to 41% smaller models, 12× faster training, 28–35% lower energy use, and +12.9% accuracy gain under non-IID data, establishing the framework as a robust and sustainable solution for IoT learning.

With the rapid expansion of next-generation networking, Internet of Things (IoT) devices have become central components of federated learning (FL) networks. FL offers a paradigm for distributed training machine learning models while preserving user data privacy. However, existing network security measures often struggle to identify legitimate contributors from opportunistic free riders within these networks. The Free Rider (FR) problem arises when participants seek to benefit from the FL processes without contributing. In particular, free riders are known to exist within or outside of the network, whereas outside free riders can hardly be identified. The Zero Trust model proposes an environment where no entity, including the network itself, is inherently trusted, providing a foundation to counter external threats seeking to exploit the network. This study proposes a novel framework strengthened by the Zero Trust model to identify external free riders in FL networks. Leveraging a Deep Autoencoding Gaussian Mixture Model (DAGMM)-based technique for internal free rider detection, our framework demonstrates superior performance in identifying free riders across various FR scenarios compared to current state-of-the-art solutions. Through our proposed framework and the principles of Zero Trust, we establish a robust security guarantee for FL networks, ensuring the integrity of the learning process.

The rapid expansion of the Internet of Things (IoT) and network services has revolutionized technology, enabling numerous intelligent applications. However, this interconnected environment also introduces significant security challenges, particularly the susceptibility of federated learning (FL) systems to poisoning attacks. Such attacks compromise the integrity of the global model by injecting malicious data, leading to inaccurate predictions and potentially endangering system reliability and user safety. While traditional approaches, such as early detection and secure aggregation methods, aim to prevent the aggregation of malicious updates, they are ineffective in addressing threats within systems that have already been compromised and did not initially implement these safeguards. This gap highlights the urgent need for robust post-compromise mitigation strategies in FL security. To address this challenge, we introduce “Unlearning as a Service for Safeguarding Federated Learning” (UaaS-SFL), a novel service designed to seamlessly integrate with any FL management system to remove the impact of poisoning clients and restore the integrity of the global model. UaaS-SFL effectively unlearns the contributions of malicious clients, ensuring both model security and system reliability. Our empirical evaluations, conducted in a simulated IoT environment, demonstrate that our service maintains model accuracy with less than a 10% deviation from the baseline achieved through retraining from scratch, underscoring the efficacy of our methodology in safeguarding FL systems. These results highlight UaaS-SFL as a critical service for securing FL management systems, providing a robust foundation for the continued growth of secure and intelligent IoT applications.

Over the past few years, the improved detection and processing features of Internet-of-Things (IoT) devices have opened the doors to several mobile crowdsourcing applications. Federated Learning (FL) is being seen as an attractive framework to address the data privacy concerns of mobile users in the context of crowdsourcing. In FL on a crowdsourcing platform, constructing an effective deep neural network (DNN) is challenging. This is primarily because the quality of the global model depends on the local model quality, which can vary greatly due to differences in the computational resources, data quantity, and data quality provided by each worker. To address these challenges, we propose QARMA-FL: Quality-aware robust model aggregation for federated learning in crowdsourcing applications, where we select the local model for aggregation based on its quality and performance. We also propose a model-quality-aware incentive mechanism to reward workers, based on their contribution to model training. Our model selection and incentive mechanism is capable of detecting Free Rider attacks, identifying workers who benefit from others contributions without contributing themselves. Most existing evaluations of FL in mobile crowdsourcing studies are not based on the real-world FL scenarios. Therefore, we evaluate QARMA-FL alongside a baseline FL model in a quantity-skew, non-IID data setup where different workers contribute varying amounts of data for model training. Our diverse experiments validated QARMA-FLs performance, demonstrating its ability to efficiently aggregate models in mobile crowdsourcing scenarios, reaching baseline results with a reduced worker participation by 40% to 60%.