Electronic healthcare (e-health) allows smart devices and medical institutions to collaboratively collect patients' data, which is trained by artificial intelligence (AI) technologies to help doctors make diagnosis. By allowing multiple devices to train models collaboratively, federated learning is a promising solution to address the communication and privacy issues in e-health. However, applying federated learning in e-health faces many challenges. First, medical data are both horizontally and vertically partitioned. Since single horizontal federated learning (HFL) or vertical federated learning (VFL) techniques cannot deal with both types of data partitioning, directly applying them may consume excessive communication cost due to transmitting a part of raw data when requiring high modeling accuracy. Second, a naive combination of HFL and VFL has limitations including low training efficiency, unsound convergence analysis, and lack of parameter tuning strategies. In this article, we provide a thorough study on an effective integration of HFL and VFL, to achieve communication efficiency and overcome the above limitations when data are both horizontally and vertically partitioned. Specifically, we propose a hybrid federated learning framework with one intermediate result exchange and two aggregation phases. Based on this framework, we develop a hybrid stochastic gradient descent (HSGD) algorithm to train models. Then, we theoretically analyze the convergence upper bound of the proposed algorithm. Using the convergence results, we design adaptive strategies to adjust the training parameters and shrink the size of transmitted data. The experimental results validate that the proposed HSGD algorithm can achieve the desired accuracy while reducing communication cost, and they also verify the effectiveness of the adaptive strategies.

Federated learning enables decentralized data own-ers to collaborate and train models in a distributed manner. A special type is Vertical Federated Learning (VFL), where each of the participated data owners only has a portion of the data features. To maintain a high accuracy and reasonable computational cost, selecting a set of features among the entire dataset is essential. Although some existing work selects features by calculating their individual contributions to the learning outcomes, knowing the joint contribution from multiple features becomes necessary but challenging. Meanwhile, security concerns are raised when calculating the joint contribution of a set of features where the feature data are stored by different owners. Using homomorphic encryption or secure computing over en-crypted data is possible, but it may cost too much when complex calculations are involved and repeated. To this end, this paper proposes a privacy-preserving feature selection protocol that considers the interactions between features stored across different data owners. Specifically, we first propose an interaction-based feature selection algorithm for vertically distributed datasets. This algorithm estimates the features' joint contributions to the model training outcomes. Then, we propose a privacy-preservation protocol to prevent the semi-honest cloud server from obtaining or inferring the raw data when aggregating the knowledge and calculating the complex interaction measure for feature selection. We create a new approximation method for interaction measures to address the high computational cost when securely calculating the interaction measure while maintaining the training accuracy. The security discussions show that the proposed protocol preserves data owner's privacy. The extensive simulations validate the achieved training accuracy and efficiency.

Digital twins represent users in the cyber world and interact between users and network controllers to better manage the mobile network. Due to communications and other resource constraints, transmitting raw data for a traditional, centralized machine learning in the mobile network has been replaced by federated learning. Federated learning allows participants to train a complex model in a distributed manner, through a group of participants' local training and a global aggregation with model updates as feedback. Although federated learning can save communications costs, address data heterogeneity and protect privacy by stopping the raw data transmission, it faces various se-curity challenges. For example, poisoning attacks may inject false models or modify existing model parameters to bias the gradient descent of federated learning. Some literature attempted to detect poisoning attacks, but the attackers can still strengthen their power by creating many identities to build their group advantage, which overturns the existing detection. In this paper, we propose a digital-twin-based Sybil detection by creating new community detection among participants in federated learning. Specifically, we first identify Sybil attackers on several levels according to their attacking strength and strategies. Then, we integrate digital twins as a side channel to distinguish Sybil identities which in fact belong to the same attacker. This could leak the attacker's correlated behavior patterns which are automatically recorded in digital twins. Under this observation, we build a DT-graph that tightly connects Sybil-controlled identities belonging to the same attacker. We propose a graph-based community detection algorithm to further partition the DT-graph and distinguish Sybil attacks. Extensive simulations validate our proposed method compared with existing work.

Mobile social networks allow users to acquire multimedia contents to their mobile devices via wireless communications. To alleviate the network traffic and latency for transmitting contents, user preferences can be predicted and popular contents can be cached at the edge of network. However, for edge caching over the indoor mobile social networks raises challenging issues. The user preference for mobile data is location-dependent in different areas of indoor environment, such that various edge nodes need to maintain its distinctive prediction model instead of using the universal one. The limited computing power for edge nodes over indoor mobile social networks also hinders effective model training on the edge. In this paper, we propose a collaborative edge caching framework that enables personalized modeling for content popularity prediction. Specifically, a non-additive measure based feature selection scheme is proposed to realize efficient yet accurate modeling on resource-constrained edge nodes. A collaborative learning algorithm is designed to reduce the computational overheads over mobile social networks by extracting global knowledge on simplifying model training through feature selection. Extensive simulation validates the effectiveness and efficiency of our proposed framework.

E-health systems consist of intelligent devices, medical institutions, edge nodes, and cloud servers to improve healthcare service quality and efficiency. In e-health systems, patients’ data are cooperatively collected by their wearable devices and the hospital they have visited, i.e., vertically distributed data. The data on wearable devices share the same feature set but are different in sample spaces, i.e., horizontally partitioned data. Meanwhile, hospitals target various user groups resulting in high data diversity, i.e., non-identically distributed data. These three characteristics cause that existing federated learning frameworks cannot efficiently train models on medical data. Furthermore, model training in e-health is time-sensitive because some diseases mutate very quickly and spread easily, which requires fast convergence of machine learning algorithms. In this paper, we address the problem of how to efficiently and rapidly train global models on e-health data. Specifically, we propose a multilayer federated learning framework to cope with data that are vertically, horizontally, and non-identically distributed. Moreover, we develop a Multi-Layer Stochastic Gradient Descent (MLSGD) algorithm towards the proposed framework to learn the optimal global model. To improve training efficiency, partial models learned by devices are aggregated on edge nodes before exchanging intermediate results with hospitals. The weight of local models is proportional to local data size when performing global aggregation to balance the impact of local models on the global model. We also prove the convergence of the MLSGD algorithm from a theoretical perspective. The experimental results from the real-world dataset MIMIC-III validate that the proposed algorithm converges fast and achieves desired accuracy.

Edge-assisted Internet of Agriculture Things (Edge-IoAT) connects massive smart devices managed by edge nodes to collect crop data for distributed computing, such as federated learning, to guide agricultural production. In Edge-IoAT, data are cooperatively collected by edge nodes and the server, i.e., vertically partitioned. In addition, sample size and distribution are different for edge nodes, i.e., horizontally partitioned. Existing federated learning frameworks are not applicable for Edge-IoAT because they do not consider both types of data partitioning simultaneously. Moreover, the excessive energy consumption may cause premature interruption of model training, and spectrum scarcity prevents a portion of edge nodes from communicating with the server. Given limited energy and communication resources, training accuracy relies on how to schedule devices. In this paper, we first propose a joint federated learning framework for Edge-IoAT to cope with both vertically and horizontally partitioned data. After that, we formulate an energy-aware device scheduling problem to assign communication resources to the optimal edge node subset for minimizing the global loss function. Then, we develop a greedy algorithm to find the optimal solution. Experiments in a Nebraska farm show that the proposed framework with energy-aware device scheduling achieves a fast convergence rate, low communication cost, and high modeling accuracy under resource constraints.

Internet of Battlefield Things (IoBT) connects massive tactical devices to collect battlefield situations and share perceived information. The IoBT can enhance the intelligent battlefield command, collaborative attack, and other applications, such as landmine trigger and post-war clearance. Existing routing path selection methods designed for wireless sensor networks (WSNs) are effective but still face challenges in IoBT scenarios. First, tactical devices follow nonuniform distributions with high density on boundaries in IoBT to prevent the location of devices from being speculated and protect strategic positions, which results in unbalanced energy consumption. Second, increasing latency in IoBT is caused by various data generation probabilities of tactical devices. Third, the military task features, such as landmine explosion, disconnection, and failure of tactical devices, may put forward special requirements on network robustness. To this end, we propose a routing path selection method with joint optimization in IoBT based on nonuniform node distributions and location-related data generation probabilities. Specifically, we first investigate and formulate the distribution and data generation probability of tactical devices. Based on the special features, energy consumption, latency, and network robustness are analyzed during multihop communications in IoBT. Then, a joint optimization problem is formulated to minimize energy consumption and latency, while maximizing the network robustness simultaneously. Furthermore, two path assignment algorithms are developed to solve this optimization problem. Finally, our simulation results show that the proposed routing path selection method can reduce energy consumption and latency with the guaranteed robustness of IoBT.

Malicious websites attempt to install malware on user’s devices without permission, which can disrupt device operation, steal personal information, and even acquire access to the device for future attacks. Accurate detection of malicious website behaviors is crucial for network security but still faces challenges. Firstly, various types and semantics of website features are required to identity the wide range of malicious characteristics, leading to massive training data and computational overhead. Secondly, to reduce model dimensionality, a proper selection of website features is essential but difficult due to the complex relations among features that can affect each other’s contribution to detection outcomes. In this paper, we propose a lightweight feature-based detection scheme against malicious websites considering the interaction measures among features and the overhead-accuracy tradeoff. Specifically, we systematically characterize the interactions among website features in a non-additive manner to indicate the aggregated impacts of feature subsets. Then we propose a quantification method to measure the feature interactions based on multivariate regression. With this method, important features are selected to substantially reduce the model dimension and computational complexity while maintaining desirable accuracy. Meanwhile, the proposed scheme provides an interpretable model that preserves the physical meanings of original features. It allows users to balance the overhead-accuracy tradeoff for detection model training through feature subset selection to fit the requirements and constraints of real applications.

With the upcoming next generation wireless network, vehicles are expected to be empowered by artificial intelligence (AI). By connecting vehicles and cloud server via wireless communication, federated learning (FL) allows vehicles to collaboratively train deep learning models to support intelligent services, such as autonomous driving. However, the large number of vehicles and increasing size of model parameters bring challenges to FL-empowered connected vehicles. Since communication bandwidth is insufficient to upload full-precision local models from numerous vehicles, model compression is usually conducted to reduce transmitted data size. Nevertheless, conventional model compression methods may not be practical for resource-constrained vehicles due to the increasing computational overhead for FL training. The overhead for downloading global model can also be omitted by existing methods since they are originally designed for centralized learning instead of FL. In this paper, we propose a ternary quantization based model compression method on communication-efficient FL for resource-constrained connected vehicles. Specifically, we firstly propose a ternary quantization based local model training algorithm that optimizes quantization factors and parameters simultaneously. Then, we design a communication-efficient FL approach that reduces overhead for both upstream and downstream communications. Finally, simulation results validate that the proposed method demands the lowest communication and computational overheads for FL training, while maintaining desired model accuracy compared to existing model compression methods.

The emerging technology of cybertwin is expected to bring revolutionary benefits to the sixth-generation (6G) network in respect of communication, resources allocation, and digital asset management. Empowered by ubiquitous artificial intelligence (AI), cybertwin is capable of adjusting the requests for computing resources to support network services by analyzing user’s demands for quality of experience and resource scarcity in the market. For resource-constrained applications, such as connected vehicles in the 6G network, cybertwin can intelligently determine the time-varying requests of computing resources for various vehicles at different times. However, the current service architecture executes AI algorithms with universal configurations for all vehicles. This causes the difficulty of customizing the complexity of AI algorithms to maintain adaptive to cybertwin’s decisions on dynamic resources allocation. In this article, we propose an adaptive AI framework based on efficient feature selection to cooperate with cybertwin’s resource allocation. This proposed framework can adaptively customizing AI model complexity with available computing resources. Specifically, we systematically characterize the aggregated impacts of all feature combinations on the modeling outcomes of AI algorithms. By utilizing nonadditive measures, the interactions among features can be quantified to indicate their contributions to the modeling process. Then, we propose an efficient algorithm to obtain accurate interaction measures for adaptive feature selection to balance the tradeoff between modeling accuracy and computational overhead. Finally, extensive simulations are conducted to validate that our proposed framework substantially reduces the overhead of AI algorithms while guaranteeing desired modeling accuracy for cybertwin-driven connected vehicles in 6G.