Network Security Management in Heterogeneous Networks

Heterogeneous networks, as a critical component of modern communication technology, have experienced rapid development in recent years [1]. The emergence of technologies like 5G [2], the Internet of Things (IoT) [3,4], and edge computing [5] has significantly enhanced the diversity and complexity of heterogeneous networks [6], making them pivotal for diverse application demands.

However, the openness and diverse characteristics of heterogeneous networks expose them to serious security challenges [7,8]. Such networks are vulnerable to attacks like Distributed Denial of Service (DDoS) attacks [9], malware propagation [10], and jamming attacks [11], posing significant risks to system stability and data privacy.

To address these pressing security challenges, researchers have developed a variety of defense strategies aimed at mitigating risks in heterogeneous networks [12,13,14,15,16,17]. Compared to traditional approaches, these strategies exhibit several distinct advantages, such as the ability to efficiently handle large amounts of data while ensuring data security, flexibility in tackling various security challenges, and resilience against advanced and persistent cyberattacks. These approaches provide significant theoretical and practical support for improving the security of heterogeneous networks.

The rapid growth of deepfake technology represents a societal risk [18]. The first contribution to this Special Issue (Contribution 1) proposes a forensic defense method with robust pseudo-Zernike moment watermarks. It employs an adaptive strategy to embed a watermark in the image background, acting as a detection marker. After experimental validation, it was found that the method can effectively improve the robustness of face-switching detection in complex environments and in the presence of disturbances. On the other hand, poor-quality images tend to hamper accurate threat detection, which, in turn, affects the proper functioning of security measures [19]. In Contribution 2, the authors propose an unsupervised low-light image enhancement method using a U-net neural network based on Retinex theory and a Convolutional Block Attention Module (CBAM). The method effectively enhances image details during image decomposition using Retinex theory, while a local adaptive enhancement function is applied to improve reflection map brightness. In addition, the designed loss function addresses the challenges of denoising, brightness enhancement, illumination smoothness, and color restoration.

The growth of low-carbon transport, spurred by environmental policies and technological advances, highlights the importance of energy trading [20]. The issue of energy trading in the electric vehicle (EV) market is beginning to attract attention from researchers. Contribution 3 presents a multi-agent reinforcement learning (MADRL)-based auction algorithm to optimize distributed energy trading in EV charger-sharing networks, enhancing social welfare and efficiency. This approach leverages blockchain technology to ensure the transparency and immutability of transactions, providing users with a transparent and decentralized trading platform. Deep reinforcement learning (DRL) techniques are used to train agents (such as EVs or charging stations) to make optimal decisions in uncertain environments. Contribution 4 presents a blockchain-based framework for secure electricity trading between EV operators and the electricity market. A Stackelberg game model using the Tiny DRL algorithm is proposed to optimize trading strategies, enhancing efficiency in uncertain markets.

Entity and relation extraction plays a key role in real-time cybersecurity monitoring and analysis [21]. Contribution 5 introduces a model for entity and relation extraction using an attention mechanism and a Graph Convolutional Network (GCN). The method uses sequence labeling for entity span detection and employs multi-feature fusion to identify all entity spans and build an entity span matrix. Next, based on the attention mechanism, the authors construct an entity relation matrix to represent correlations between entities. Finally, the entity span matrix and entity relation weighted matrix are fed into the GCN for unified entity and relation extraction.

Contribution 6 addresses the challenge of drug repositioning using Graph Neural Networks (GNNs) to model complex relationships between drugs, diseases, and their subcategories. It incorporates a prototype-based feature-enhancement mechanism (PFEM) along with a dual-task classification head (D3TC) to enhance the representation of these relationships. The proposed method was experimentally validated on four public datasets. The results showed that the method surpasses state-of-the-art approaches, significantly enhancing drug repositioning accuracy and efficiency.

The rapid growth of the Internet of Vehicles (IoV) has led to increased focus on its security challenges [22]. Federated learning (FL) can address security and privacy concerns in the Intelligent Connected Vehicle (ICV) domain [23,24], but still faces challenges like multimodal data integration [25], Byzantine attacks [26], and communication limits [27]. In response to these challenges, Contribution 7 introduces a Byzantine-robust multimodal FL framework to tackle these issues. It counters Byzantine attacks with a gradient compression-based aggregation technique. It incorporates a multimodal learning framework to improve adaptability to complex environments and uses top-k gradient compression to enhance communication efficiency. Contribution 8 addresses the privacy leakage risks associated with model parameter exchange in the peer-to-peer (P2P) architecture of FL for IoV scenarios. To address these risks, a differential privacy scheme is proposed, allowing nodes to dynamically adjust noise levels in their model parameters based on their distances to other nodes. This method balances security and model quality.

In edge computing, security is a key factor affecting system performance and user trust [28,29,30]. The issue of security in Mobile Edge Computing (MEC) scenarios remains an open challenge [31]. To address this, Contribution 9 presents a DRL-based security-aware task-offloading framework. This framework uses an Advanced Encryption Standard (AES) to secure data during task offloading. Furthermore, the task-offloading process is modeled as a Markov Decision Process (MDP) and optimized with a Proximal Policy Optimization (PPO) algorithm to reduce latency and energy use. Contribution 10 introduces an approach based on a large language model (LLM) called the Spatio-Temporal Large Language Model with Edge Computing Servers (STLLM-ECS) to predict industrial production nationwide PM_2.5. To address security risks in centralized training, such as data leaks during transmission, the paper introduces an edge-distributed learning framework, STLLM-ECS. The framework uses a novel method, called NodeSort, to divide the nationwide sensor network graph into subgraphs. Data and training tasks for each subgraph are allocated to separate Edge Computing Servers (ECSs), reducing data leakage risks.

To address security challenges in heterogeneous networks, Contribution 11 introduces the optimized multi-objective multipath transmission algorithm (MOMTA-HN). This algorithm integrates multiple objectives into path selection, allowing the calculation of optimal paths. Using these redundant paths, the algorithm provides enhanced protection for communication processes within heterogeneous networks.

Contribution 12 presents MedMixtral 8x7B, a medical LLM using the mixture-of-experts (MoE) architecture and an offloading strategy for IoMT deployment. Using the proposed efficient inference-offloading strategy, the model dynamically allocates its weights between the CPU RAM and disk during run-time, effectively reducing GPU memory consumption. This approach enables the deployment of MedMixtral 8x7B on resource-constrained IoMT devices, thus enhancing user privacy protection.

Community detection is a crucial method for analyzing complex systems and organizational structures [32]. Contribution 13 reinterprets community structure by encoding edge information, highlighting its essence by reducing transmitting edge information uncertainty within community structures. This new definition better captures the intrinsic characteristics of communities. Based on this concept, the community detection algorithm CSIM is proposed, which aims to maximize community structure information as its optimization objective and efficiently approximates optimal community partitioning, with its practical effectiveness validated through experiments.

Scrap detection is key to linking the smelting process with the industrial internet, prioritizing security and privacy [33]. Contribution 14 introduces FedScrap, a layer-wise personalized FL framework that coordinates decentralized scrap data while safeguarding privacy. FedScrap uses a self-attention mechanism to aggregate personalized client models layer-by-layer, prioritizing data relevance. This approach enhances the accuracy of model aggregation while addressing data heterogeneity and protecting data privacy.