This paper addresses the challenges of privacy protection, device constraints, resource heterogeneity, and trusted inference in the integration of large language models (LLMs) with Internet of Things (IoT) devices, proposing a Hierarchical Federated Collaborative Computing (HFCC) framework. Unlike traditional federated learning, HFCC employs horizontal splitting + chunking to reduce LLMs computation overhead. The framework dynamically splits LLMs into: 1) global shared layers optimized by edge servers, and 2) device-local layers trained on private data, ensuring raw data remains on-device. During inference, chunking of shared layers and dynamic task allocation adjust computational loads based on real-time device states, mitigating high-load security risks. Furthermore, leveraging the hierarchical federated learning architecture, the system employs an anonymized parameter aggregation mechanism during training to achieve multi-level privacy protection. Simultaneously, a cross-device consensus verification mechanism performs trusted validation of distributed intermediate results, effectively identifying malicious node behavior. Experiments show 58% faster inference in resource-constrained environments, significantly reduced data exposure risks, and 94% malicious node detection accuracy versus traditional federated learning. This lays a solid foundation for building an intelligent, efficient, and secure IoT ecosystem.

In the context of 6G communication networks, large communication models emerge as the core application for edge communication networks. They hold immense application potential in areas like intelligent connected vehicles, collaborative robots, and holographic communications. However, practical implementations of these applications face challenges such as unstable learning environments, resource constraints, information theft, and heterogeneity of edge devices. This paper delves into the distributed deployment of large communication models under resource-constrained nodes and introduces a distributed architecture based on cloud-edge collaboration. Crucially, we introduce a negotiation and debate mechanism between small AI models. This mechanism adopts innovative strategies, enabling it to match or even surpass the capabilities of large language models in terms of inference and accuracy. Furthermore, addressing the data confidentiality issues of large model parameter transmissions, this study proposes a secure federated learning mechanism based on privacy computing. Combining novel model layering technology with privacy protection measures, it aims to enhance the security of model parameter transmissions. Experimental results confirm that while improving learning efficiency and inference accuracy, this strategy successfully ensures the privacy security of edge nodes, offering a practical solution for future privacy protection.

As maritime communication networks become more developed and widely used, an increasing number of end devices are being connected to these networks. Edge computing is an effective way to meet the realtime computing needs of these devices. However, with federated learning becoming more prevalent in edge computing decision-making, resource-constrained end devices are becoming more vulnerable to security threats. In particular, poisoning attacks are a significant security concern in the federated learning training process as local data is invisible to the outside world, enabling malicious participants to easily tamper with it. To address this issue, this paper proposes a Common Layer Mean Detection method for poisoning attacks that reduces their impact on the accuracy and resource consumption of the federated model while ensuring its security. The proposed method identifies poisoning attacks by comparing the mean distribution differences between attackers and honest clients in the common layer. Aggregation weights are then set based on the detection results to eliminate the impact of spurious parameters of malicious participants on the overall model. The effectiveness of this approach is demonstrated by comparing it with related schemes in terms of security, communication overhead, and computational overhead. Overall, the proposed method is shown to be both secure and efficient, making it a valuable addition to the field of federated learning for maritime communication networks.

With the increasing demand for high performance and low power consumption in embedded real-time systems, performance asymmetric multiprocessor architecture is beginning to be applied to embedded real-time systems, because this heterogeneous architecture allows the system to allocate computational resources as needed to satisfy the demands of each application and dynamic workloads. Most of the previous studies on EDZL scheduling algorithms have focused on task prioritization, and they simply assign the highest priority tasks to the fastest processors. In this paper, we choose to assign the highest priority tasks on the slowest processor and propose SSFEDZL (slowest speed fit for earliest deadline first until zero-laxity) scheduling algorithm. We compare SSF-EDZL scheduling algorithm with other scheduling algorithms and the experimental results show that SSF-EDZL is a better scheduling algorithm which has higher processor utilization and better scheduling performance.

The emergence of digital twin technology is expected to reduce the significant cost of traditional physical debugging. Moreover, it can also fully combine IoT real-time data, large data analysis, and simulation, to make the best decisions. In the digital twin network, the data distribution and computing power of different device nodes shows great heterogeneity, but the internal node data are independent and identically distributed. Therefore, our learning algorithm should be able to learn not only the network commonality, but also the node specificity. In order to provide secure and personalised services for different device nodes, a polymorphic learning (PL) framework is proposed in this paper. PL divides the classical neural network model into a homomorphic model and a polymorphic model to realise flexible network control. Finally, the advantages of the PL algorithm in the collaborative optimisation of different nodes are proven through simulation experiments, and the algorithm running process is simulated through a classic case, with the final results proving the superiority of the algorithm.