Deep neural networks (DNNs) deployed on edge devices often suffer from severe performance degradation when exposed to dynamic and continually shifting environments. Test-time adaptation (TTA) has emerged as a promising solution by updating models online with incoming test data. However, edge deployment poses unique challenges: limited computational resources, latency caused by adaptation delays, and knowledge isolation across devices. The situation becomes even more complex in open-world scenarios, where the presence of unknown categories further disrupts adaptation. To overcome these limitations, we propose TeamTTA, a cloud-integrated framework designed for efficient multi-device collaboration open-set test-time adaptation. Specifically, TeamTTA aggregates reliable samples from multiple edge devices through crowdsourcing, uploads them to the cloud, and maintains a memory buffer for continual adaptation. A large vision model (LVM) in the cloud leverages its zero-shot generalization ability to filter out open-set samples and acts as a teacher model, distilling its knowledge into a replicated student edge model stored in the cloud. The adapted model parameters, or alternatively global statistics under poor network conditions, are then transmitted back to the edge devices for efficient inference. Extensive experiments on standard public TTA benchmarks, including corrupted and open-set datasets, show that TeamTTA achieves superior adaptation accuracy, robustness to distribution shifts, and communication efficiency, outperforming state-of-the-art TTA baselines. These results validate the effectiveness of integrating cloud-edge collaboration and LVM-driven knowledge distillation for real-world edge intelligence.

The Vision Transformer (ViT) excels in accuracy when handling high-resolution images, yet it confronts the challenge of significant spatial redundancy, leading to increased computational and memory requirements. To address this, we present the Localization and Focus Vision Transformer (LF-ViT). This model operates by strategically curtailing computational demands without impinging on performance. In the Localization phase, a reduced-resolution image is processed; if a definitive prediction remains elusive, our pioneering Neighborhood Global Class Attention (NGCA) mechanism is triggered, effectively identifying and spotlighting class-discriminative regions based on initial findings. Subsequently, in the Focus phase, this designated region is used from the original image to enhance recognition. Uniquely, LF-ViT employs consistent parameters across both phases, ensuring seamless end-to-end optimization. Our empirical tests affirm LF-ViT's prowess: it remarkably decreases Deit-S's FLOPs by 63% and concurrently amplifies throughput twofold. Code of this project is at https://github.com/edgeai1/LF-ViT.git.

Photographs taken in low-illumination environment have a low signal-to-noise ratio and impaired visual quality. Enhancing lowlight images tends to amplify noise. To address this problem, we propose a Multi-Scale Recursive Feature Enhancement (MSRFE) network for low light image enhancement. The MSRFE network consists of several Feature Enhancement (FE) blocks which are applied to enhance the multi-scale image feature and remove the noise recursively in each scale residual map between adjacent scale feature. Then, a deep recursive Curve Adjustment (CA) block is proposed further fine-tunes the output of MSRFE netowrk by learning a non-linear curve which can adjust the image luminance and details. We evaluate the proposed method on both real and synthetic datasets. The results show that our proposed method outperforms other state-of-the-art methods on both visual and objective evaluation indicators.

Offloading task to roadside units (RSUs) provides a promising solution for enhancing the real-time data processing capacity and reducing energy consumption of vehicles in the vehicular ad-hoc network (VANET). Recently, multi-agent deep reinforcement learning (MADRL)-based offloading approaches have been widely used for task offloading in VANET. However, existing MADRL-based approaches suffer from offloading preference inference (OPI) attack, which utilizes the vulnerability in the policy learning process of MADRL to mislead vehicles to offload tasks to malicious RSUs. In this paper, we first formulate a joint optimization of offloading action and transmitting power with the objective of minimizing the system cost, including local and edge costs, under the privacy requirement of protecting offloading preference during offloading policy learning process in VANET. Despite the non-convexity and centralized of this joint optimization problem, we propose a privacy-aware MADRL (PA-MADRL) approach to solve it, which can allow the offload decision of each vehicle to reach the Nash Equilibrium (NE) without leaking offloading preference. The key to resisting the OPI attack is to protect the offloading preference by 1)elaborately constructing the noise based on ($\beta,\Phi $)-differential privacy mechanism and 2) adding it to the action selection and policy updating process of vanilla MADRL. We conduct a detailed theoretical analysis of the convergence and privacy guarantee of the proposed PA-MADRL, and extensive simulations are conducted to demonstrate the effectiveness, privacy-protecting capacity, and cost-efficiency of PA-MADRL approach.

Industrial Internet of Things (IIoT) devices are widely used for monitoring and controlling the process of automated manufacturing. Owing to the limited computing capacity of the IIoT sensors in the production line, the scheduling task in the production line needs to be offloaded to the edge computing server (ECS). To obtain the desired quality of service (QoS) during offloading scheduling tasks, the precise interaction information between the production line and ECSs has to be uploaded to the cloud platform, which poses privacy issues. The existing works mostly assume that all the interaction information, i.e., the offloading decision for the subtask in a scheduling task, has same privacy level, which cannot meet the various privacy requirements of the offloading decision for the subtask. Hence, we propose a local-differential-privacy-based deep reinforcement learning (LDP-DRL) approach in the edge-cloud-assisted IIoT to provide personalized privacy guarantee. The LDP mechanism can generate different levels of noise to satisfy the various privacy requirements of the offloading decision for the subtask. The prioritized experience replay is integrated in DRL to reduce the impact of noise on the QoS performance of task offloading. The formal analysis of LDP-DRL is provided in terms of privacy level and convergence. Finally, extensive experiments are conducted to evaluate the effectiveness, the capacity of privacy protection, the impact of discount factor on the convergence, and the cost efficiency of the LDP-DRL approach.

The high computational capacity demanded in blockchain mining hinders the involvement of mobile devices due to their limited computation power. Offloading the blockchain mining task to base stations (BSs) is a promising solution for mobile blockchain. Recently, many reinforcement learning (RL)-based approaches achieve the long-term performance of Quality of Service (QoS) during online computation offloading, but they fail to consider the risk of privacy leakage. Existing works for privacy preserving share an unresolved problem that they provide a private mechanism by adding private constraints into the value function of RL algorithm but neglect to protect the value function itself. Hence, we investigate a novel privacy issue caused by value function leakage, named offloading preference leakage. To solve this issue, we propose a privacy-aware deep RL method (PA-DRL) for computation offloading over the mobile blockchain. Specifically, a functional noise is generated, then added to the exploring and policy updating processes of DRL. Furthermore, we adopt a cooperative exploring mechanism and prioritized experience replay (PER) to improve the convergence rate of the proposed method. We provide the theoretical analysis for privacy preserving and convergence. Finally, simulation results show that our method can perform cost-efficient computation offloading, compared with benchmark methods.

Unmanned Aerial Vehicles (UAV) have rapidly developed and been widely applied to military and civilian applications in recent years. Anomaly Detections and finding out the root causes are critically important for UAV social network security. In the UAV social networks, the drone can communicate with one another directly in a form of leading flights with followers during a far away mission. The ground controller cannot get their information directly. Besides, none of the works consider the privacy protection and anomaly root cause tracing during the distributed detection. This paper presents a self-verification approach among UAV flights which can check whether the flights have honestly obeyed the orders or suffered the anomalies. Besides, we do the verification without looking through the plaintext records or data of the drones. Finally, to instruct the drones to solve the problems, we trace the fundamental root causes leading to the anomalies by learning the fault tree. We apply our approach on raw UAV social network data and align our experiment with two former works as baselines for comparison. Our approach can reduce the time cost of verification from exponential growth to linear growth and improve the tracing accuracy rate around 4.3% higher than the former work.

Unmanned Aerial Vehicle (UAV) has enormous potential in many domains. According to the characteristics of UAV, it is important for UAV network to assure low latency and integrity and authentication of commands sent by command center or command stations to UAV. In this paper, we proposed a UAV network architecture based on mobile edge computing (MEC) which helps guarantee low latency in the UAV network. Afterwards, we proposed a designated verifier proxy blind signature (DVPBS) scheme for UAV network and proved that it is existentially unforgeable under an adaptive chosen message attack in the random oracle model. We compared the efficiency of our DVPBS scheme with other signature schemes by implementing them in jPBC and theoretically analyzing their signature length. The experiment results indicate that our DVPBS scheme is efficient. The signature length of our DVPBS is longer, but it is still short enough compared with the transmission capacity of UAV.

Cyberspace is the digital world created based on traditional physical, social, and thinking spaces (PST) but in turn makes a great difference on PST. The cyberization and the emergence of cyber-enabled spaces can be viewed as the bridge between cyberspace and PST, which reshaped the current definition of cyberspace and contributed to a novel concept general cyberspace (GC). Generally, GC is a unified description of conventional cyberspace (also shortly cyberspace in this paper) and cyber-enabled PST. It essentially emerges from cyberspace based on ubiquitous connections between things and the deep convergence of spaces. This paper proposes the definition of GC and investigates it from its three main aspects: 1) existence; 2) interactions; and 3) applications/services, respectively, in terms of philosophy, science, and technology outlook.

Cyberspace is a new basic space after the three traditional basic spaces-physical, social, and thinking spaces (PST spaces). It is a trend that entities (objects) and PST spaces they are living in to be cyberized. On the one hand, the rapidly developing of cyberspace has the increasingly significant influences to PST spaces. On the other hand, the cyberization of objects in PST spaces have been continuously deepening and strengthening. Cyberization leads to the convergence of the four basic spaces, which also called cyberspace and cyber-enabled physical-social-thinking spaces (CPST spaces). In recent years, the philosophy research on CPST spaces and objects (short for cyber philosophy) has been developing rapidly while some researchers try to figure cyber science and its fundamental issues. Up to now, the bridge, fundament logic from cyber philosophy to cyber science, has not yet formed. This paper proposes a new concept of “cyberlogic” for establishing a bridge from cyber philosophy to cyber science. The etymology, concept, contents, and methods of cyberlogic are presented, and the cyberlogic for the CPST spaces is shown. Moreover, main issues and methodologies for cyberlogic are discussed.