Information leakage and model attacks pose risks in the analysis and exchange of medical data, as the language models used to process medical records may retain training data. Traditional models, on the other hand, are often too complicated and use old ineffectual methods for removing personal data. This can compromise the data’s integrity and quality, making it less useful for future tasks, especially when combined with other language models. This paper introduces the Self-Decoded Model of Medical De-identification (SDM-M-DID). The model employs a secure BERT-based encoder to paraphrase sensitive data, ensuring HIPAA compliance. Unlike traditional models that only mask sensitive tokens, the SDM-M-DID decodes its own embeddings to generate an internal representations of these tokens. Then, it integrates this representations with the pre-trained BERT dictionary to rephrase tokens, preserving their semantic role while altering grammar to prevent re-identification. Compared to existing large language models, our model achieves a score of 0.8416 F1 BERTscore, striking an optimal balance between the variability and similarity of deidentified tokens. We conducted experiments on two medical datasets to demonstrate the effectiveness of the model. Metrics show that there is only a ± 1 % to ± 2 % difference in accuracy between the original datasets and the de-identified datasets. In total, this demonstrates that SDM-M-DID not only effectively preserves data integrity and is not inferior in efficiency to new large language models but even improves it in some cases while using a more secure and less resource-intensive technology.

With the growing demand for seamless global connectivity, the integration of Non-Terrestrial Networks (NTNs) and terrestrial networks presents new opportunities for largescale Federated Learning (FL). To reduce the communication overhead in NTN-assisted FL systems, NTN-assisted Hierarchical Federated Learning (HFL) frameworks have been proposed, which leverage aerial platforms as intermediate aggregation layers to effectively alleviate the communication burden on remote cloud servers. However, the high orbital altitude and mobility of Low-Earth Orbit (LEO) satellites introduce additional communication latency and frequent changes in network topology, making synchronous aggregation at the edge-cloud layer significantly degrade the training efficiency of HFL. Existing hybrid asynchronous HFL frameworks typically rely on static asynchronous triggering strategies that are not well-suited for the dynamic NTN environments and overlook the model drift caused by data heterogeneity among edge nodes. Furthermore, the data and heterogeneity among clients pose further challenges in HFL training performance. To address these issues, we propose a Dynamic Hybrid Asynchronous Hierarchical Federated Learning (DHA-HFL) framework for NTN scenarios. At the edge-cloud layer, we introduce a dynamic semi-asynchronous aggregation mechanism, consisting of an adaptive asynchronous aggregation trigger mechanism and an asynchronous dual-end aggregation strategy. The former adaptively adjusts the global aggregation timing to accommodate the dynamic NTN environments, while the latter enables efficient cloud-edge coordination to mitigate global model drift induced by data heterogeneity and inherent model staleness. At the client-edge layer, we employ a synchronous aggregation mechanism and propose a heterogeneityaware adaptive local iteration control strategy. By deriving the convergence bound of synchronous FL under non-IID data, we design an iterative algorithm to optimize the local iteration count of devices, minimizing training latency at the client-edge layer and accelerating global model convergence. Extensive experiments demonstrate the superior performance of DHA-HFL in terms of training latency, model accuracy, and convergence speed, providing an efficient distributed learning solution for NTN-assisted IoT scenarios.

Leveraging flexible deployment and wide-area coverage, Unmanned Aerial Vehicles (UAVs) assisted edge computing (EC) extends computation capability to the network edge and has become a key paradigm for improving the quality of service in the Internet of Things. However, constrained by onboard resources, UAVs struggle to independently process massive heterogeneous tasks in real time. Existing studies mainly focus on cooperation between UAVs and ground devices or the cloud, while the coupled cooperative relationships among UAVs and the performance scalability in large-scale UAV networks remain insufficiently characterized. To this end, we propose a distributed task assignment framework for a multi-UAV cooperative EC system, where each UAV explicitly accounts for cooperation with other UAVs and, leveraging controllable mobility, assigns a portion of local tasks to other UAVs or base stations to minimize the long-term system-wide total latency. To address the resulting joint trajectory control and task assignment optimization problem, we first formulate it as a Markov decision process. We then propose a Multi-head self-Attention critic-assisted MADDPG (MA$^{2}$DDPG) algorithm to train local online decision models for UAVs. Under the fully cooperative setting, we employ a centralized Critic to exploit global information and reduce parameter redundancy; concurrently, a multi-head self-attention mechanism is incorporated into the Critic to aggregate multi-UAV interaction information, delineate implicit cooperative dependencies, and alleviate the network input dimensionality curse arising from increasing UAV scale. Finally, extensive simulation experiments validate the effectiveness of the proposed method.

Federated learning (FL) presents a promising strategy for distributed and privacy-preserving learning, yet struggles with performance issues in the presence of heterogeneous data distributions. Recently, a series of works based on sharpness-aware minimization (SAM) have emerged to improve local learning generality, proving to be effective in mitigating data heterogeneity effects. However, most SAM-based methods do not directly consider the global objective and require two backward pass per iteration, resulting in diminished effectiveness. To overcome these two bottlenecks, we leverage the global model trajectory to directly measure sharpness for the global objective, requiring only a single backward pass. We further propose a novel and general algorithm FedGMT to overcome data heterogeneity and the pitfalls of previous SAM-based methods. We analyze the convergence of FedGMT and conduct extensive experiments on visual and text datasets in a variety of scenarios, demonstrating that FedGMT achieves competitive accuracy with state-of-the-art FL methods while minimizing computation and communication overhead. Code is available at https://github.com/harrylee999/FL-SAM.

Highly infectious diseases like COVID-19 have affected millions worldwide, with transmission clearly linked to crowd mobility. However, research on the predictive power of city-level mobility remains limited. Most existing forecasting methods focus on state or national level, making it difficult for policymakers to quickly respond to emerging threats. To address this gap, we propose the Inner-Inter City Learner (IIL) for short- and mid-term COVID-19 case predictions at the city level, based on the correlation between human interaction and new cases. IIL consists of two key components: an inter-city transmission learner and an inner-city propagation learner. The first uses city mobility graphs and graph neural networks to learn how transmission in one city is influenced by others. The second, leveraging the highly contagious nature of the virus, captures key features of COVID-19 spread within cities. To overcome limited inner-city mobility data, we apply model-agnostic meta-learning to transfer common features across cities. We conduct various experiments and compare our methods with the state-of-art baselines. The results show the superiority of our method across various forecast horizons.

Lossless floating-point time series compression is crucial for a wide range of critical scenarios, such as data transmission in the Internet of Things. In this paper, we propose a lossless floating-point compression method Elf. that employs a set of optimizations for the encodings of significand counts, leading zeros, trailing zeros and sharing conditions. Specifically, we first devise a Huffman-based method for the significand counts along with erasing flags. Then, we develop a dynamic programming algorithm with a set of pruning strategies to efficiently compute the adaptive approximation rules for leading zeros and trailing zeros, respectively. Next, we propose an adaptive sharing condition for the counts of leading zeros and trailing zeros. We further extend Elf. to Streaming Elf., i.e., SElf., which achieves almost the same compression ratio as Elf., while enjoying higher efficiency. We compare Elf. and SElf. with 9 competitors using 14 datasets, demonstrating the powerful performance of both Elf* and SElf*. All the source codes and datasets are publicly released.

With the prosperity of deep learning (DL) in Internet of Things (IoT) fields, federated learning (FL), viewed as a critical element of numerous DL-aided IoT intelligent applications, enables cooperative DL training across decentralized clients without revealing their personal data. However, the computational capacity heterogeneity and limited energy resources of IoT devices cause a huge negative influence on FL training in IoT intelligence applications. To address the above issues, this paper proposes an excellent distributed training mechanism for hierarchical FL to reduce training latency and energy cost for achieving the desirable accuracy. Specifically, taking into account computational capacity heterogeneity of clients, we first design a latency-regularization-aware client selection algorithm to appropriately select participating clients in training epochs and control their participation frequencies for boosting distributed training efficiency. Subsequently, after obtaining the selected client subset in each hierarchical FL training epoch, by leveraging the variable transmission delays of clients in distributed training, we propose a mixed integer linear programming-based transmission power management strategy for participating clients to alleviate their energy consumption burden. Extensive numerical results demonstrate that our proposed mechanism can attain 576.93% training speedup and achieve 10.94% accuracy enhancement compared with the baseline General FL, and yield up to 56.91% energy cost savings compared with the baseline HierFAVG.

Federated learning participants have two identities: model trainers and model users. As model users, participants care most about the performance of the final model on their own distributions, which is called Personal Model Performance (PMP). This makes training a single global model to accommodate all participants impractical because the data distributions of participants are heterogeneous. As model trainers, due to high training costs, participants are reluctant to contribute models if incentives are not enough. With the combination of the above two reasons, we propose a dual-identity double auction as an incentive mechanism in personalized federated learning, allowing directional selection between model users and model trainers, both of which are served by FL participants. Within the double auction framework, we devise a reinforcement learning-based model selection method. This method selects a set of models for each buyer to bid on. The bought models are aggregated to be a personalized model to achieve higher PMP. Additionally, we implement a transaction partition-based approach for determining clearing prices and winning pairs. We address the challenge of the unavailability of private yet essential data distribution information, the coupled influence of model selection and auction results on PMP, and more utility improvement ways of multi-demand dual-identity participants. Finally, our double auction optimizes the PMP of all participants and ensures the truthfulness of multi-demand dual-identity participants, which is harder compared with single-demand single-identity participants.

Multi-access edge computing has become an effective paradigm to provide offloading services for computation-intensive and delay-sensitive tasks on vehicles. However, high mobility of vehicles usually incurs spatio-temporal load-imbalances among edge servers. Therefore, task migration is employed to maintain dynamic workload balancing by transmitting excessive tasks from overloaded to underloaded servers. Recent studies adopt deep reinforcement learning approaches to generate offloading and migration decisions based on current observations of systems. However, we argue that the migration direction is highly dependent on vehicular movements, and task migration towards the wrong direction could lead to additional delays. Therefore, we emphasize the importance of guiding task migration via exploring prospective trajectories of vehicles. We propose a Mobility-Aware Cooperative Multi-Agent (MCMA) deep reinforcement learning approach to make vehicle-by-vehicle decisions in multi-edge computation offloading scenarios. A two-stage decision framework is designed to solve the joint optimization problem of computation offloading and resource allocation. Additionally, an Informer-based multi-step vehicular trajectory prediction module is incorporated to enhance the capability of forecasting vehicular movements. Extensive experiments and analysis are conducted on synthetic and realistic scenarios, showing that our approach consistently outperforms both heuristic and DRL-based methods. The simulation scenarios and source codes are publicly available here.

Wireless powered mobile edge computing combines the benefits of wireless power transfer and mobile edge computing, enabling mobile devices (MDs) to overcome energy and computational constraints by obtaining energy from hybrid access point (HAP) and offloading computation tasks to HAP. However, given the devices' half-duplex nature, conflicts arise between time allocation and task offloading. Current research lacks a comprehensive examination of the joint impact of task offloading and time allocation on long-term delay. Furthermore, most existing centralized solutions may encounter issues such as dimensionality explosion and so on. This paper thoroughly investigates the task offloading problem in wireless powered mobile edge computing systems, focusing on minimizing long-term task completion delay under delay and energy constraints. To address this problem, we reformulate the task offloading process as a bi-level Markov decision process and convert it into an optimization problem of finding the optimal policy. Subsequently, we introduce a Bi-level Actor-Critic-based Task Offloading Approach (BiAC-TOA), where MDs and HAP leverage centrally trained agents for decisions on task offloading and time allocation in a decentralized manner. Finally, a series of experiments confirm the superiority of our proposed approach.

The distributed nature of multitier swarm attacks renders it more difficult for a single-tier intrusion detection system (IDS) to secure cross-tier computation offloading in multitier sensor edge cloud (SEC). To perceive and prevent such attacks, we model IDSs in different layers as an IDS federation network (IDFN) and present a generic framework to prevent cooperative attacks and reconfigure the defense strategy of IDFN across the three-tier SEC. The framework provides single-tier, two-tier, and three-tier dynamic awareness models based on the susceptible-infected-susceptible (SIS) dynamical equations to characterize the update process of message states to obtain the equilibrium solution between alarm messages and normal messages captured by IDSs. For swarm attack events from multitier SEC, we model the cross-tier cooperative interactions between IDSs and swarm attackers as an event–condition–action (ECA) regret learning game (ERLG) to achieve a distributed IDS reconfiguration to reduce the overall SEC alarm messages while ensuring the equilibrium of message states with the cooperation of IDSs. Simulation results demonstrate that our proposed scheme is superior to other reconfiguration mechanisms under swarm attacks in three-tier SEC.

Federated learning (FL) is an attractive industrial paradigm to accomplish distributed artificial intelligence (AI) training collaboratively in a data privacy-preserving manner. Most existing designs for FL systems assume that industrial user equipments (UEs) participate voluntarily in FL training. However, since both AI model training and transmission consume considerable energy, UEs are reluctant to participate without economic rewards. Hence, the lack of proper economic reward incentive mechanism results in low UE utility and frustrates UEs' enthusiasm for participating in training. To address the above challenge, in this article, we propose a two-phase energy-aware reward incentive mechanism for the edge-cloud-assisted hierarchical federated learning (HFL) system to optimize the overall UE utility, thereby, incentivizing UEs to participate more actively. Specifically, at the cloud server phase, we design an energy quantity-aware incentive mechanism for reasonably distributing rewards to its sub-edge-assisted FL systems. Subsequently, at the edge server phase, based on the quantitative analysis for the optimal reward allocation solution, we develop an energy-aware water filling-based reward incentive mechanism to adapt to individual needs of UEs and maximize the overall UE utility. Experiments verify that, compared to well-known benchmarks, our incentive mechanism can improve the overall UE utility by up to 55.94% and better incentivize UEs to participate in training.

Fully supervised semantic segmentation has performed well in many computer vision tasks. However, it is time-consuming because training a model requires a large number of pixel-level annotated samples. Few-shot segmentation has recently become a popular approach to addressing this problem, as it requires only a handful of annotated samples to generalize to new categories. However, the full utilization of limited samples remains an open problem. Thus, in this article, a mutually supervised few-shot segmentation network is proposed. First, the feature maps from intermediate convolution layers are fused to enrich the capacity of feature representation. Second, the support image and query image are combined into a bipartite graph, and the graph attention network is adopted to avoid losing spatial information and increase the number of pixels in the support image to guide the query image segmentation. Third, the attention map of the query image is used as prior information to enhance the support image segmentation, which forms a mutually supervised regime. Finally, the attention maps of the intermediate layers are fused and sent into the graph reasoning layer to infer the pixel categories. Experiments are conducted on the PASCAL VOC-$5^i$dataset and FSS-1000 dataset, and the results demonstrate the effectiveness and superior performance of our method compared with other baseline methods.

WiFi-based human gesture recognition has recently enjoyed increasing popularity in the Internet of Things (IoT) scenarios. Simultaneously recognizing user identities and user gestures is of great importance for enhancing the system security and user quality of experience (QoE). State-of-the-art approaches that perform dual tasks suffer from increased latency or degraded accuracy in cross-domain scenarios. In this paper, we present WiDual, a dual-task system that achieves accurate cross-domain gesture recognition and user identification based on WiFi in a real-time manner. The basic idea of WiDual is to use the attention mechanism to adaptively explore cross-domain features worthy of attention for dual tasks. WiDual employs a CSI (Channel Statement Information) visualization method that transfers WiFi signals to images for further feature extraction and model training. In this way, WiDual mitigates the possible loss of useful information and excessive delays caused by extracting handcrafted features directly from the WiFi signal. Furthermore, WiDual utilizes a collaboration module to combine gesture features and user identity features to enhance the performance of dual-task recognition. We implement WiDual and evaluate its performance extensively on a public dataset including 6 gestures and 6 users performed across domains. Results show that WiDual outperforms state-of-the-art approaches, with 26% and 8% improvements on the accuracy of cross-domain user identification and gesture recognition respectively.

Urban spatio-temporal (ST) prediction plays a crucial role in smart city construction. Due to the high cost of ST data collection, improving ST prediction in a lack of data is significant. For this purpose, existing meta-learning methods have been demonstrated powerful by learning an initial network from training tasks and adjusting to target tasks with limited data. However, such shared knowledge from a set of tasks may contain irrelevant noise due to the gap of region-varying ST dynamics, resulting in the negative transfer issue. As a revelation of regional functional patterns, region profiles give rise to the diversity of ST dynamics. Thus, we design a novel adaptive meta-optimized model MetaRSTP, which conducts the initial prediction model in a finer-granularity of region level with region profiles as semantic evidence. To enhance the expressiveness of profiles, we firstly build a semantic alignment space to explore the inter-view co-semantics. Fusing it with view-specific uniqueness, the multi-view region profiles can be better applied in urban tasks. Then, a regional bias generator derives non-shared parameters in terms of profiles, which alleviates the divergence among regions. We set a new meta-learning strategy as initialize the network with fixed generalizable parameters and region-adaptive bias, thus enhancing the personalized prediction performance even in few-shot scenarios. Extensive experiments on real-world datasets illustrate the effectiveness of our MetaRSTP and our learned region profiles.

As a popular location-based service, next point-of-interest (POI) recommendation predicts users’ next movements based on recent visits. Existing works mostly focus on individual sequential preference mining, ignoring the crucial crowd transition patterns under the effects of corresponding spatio-temporal (ST) contexts. Intuitively, users also follow crowd transition rules. In order to better exploit the extensive ST-specific collaborative signals among users, we propose a novel next POI recommender, including both ST-adaptive crowd-enhanced preference and user-specific multi-semantic interest modeling. We firstly introduce user-agnostic region-and timeslot-level graphs to capture ST-specific crowd transition patterns. As they imply fine-grained general preferences and function as prior meta knowledge, we employ a meta gated recurrent network to make an adaptive prediction for specific ST-context in a meta-learned unified way. Moreover, in modeling user-specific interest, we extract multi-semantic correlations from graph-augmented personal trajectories to obtain high-quality preference representation. Extensive experiments on real-world datasets show the superiority of our proposed model against seven state-of-the-art methods.

Weakly supervised semantic segmentation (WSSS) with image-level annotations has achieved great processes through class activation map (CAM). Since vanilla CAMs are hardly served as guidance to bridge the gap between full and weak supervision, recent studies explore semantic representations to make CAM fit for WSSS and demonstrate encouraging results. However, they generally exploit single-level semantics, which may hamper the model to learn a comprehensive semantic structure. Motivated by the prior that each image has multiple levels of semantics, we propose hierarchical semantic contrast (HSC) to ameliorate the above problem. It conducts semantic contrast from coarse-grained to fine-grained perspective, including ROI level, class level, and pixel level, making the model learn a better object pattern understanding. To further improve CAM quality, building upon HSC, we explore consistency regularization of cross supervision and develop momentum prototype learning to utilize abundant semantics across different images. Extensive studies manifest that our plug-and-play learning paradigm, HSC, can significantly boost CAM quality on both non-saliency-guided and saliency-guided baselines, and establish new state-of-the-art WSSS performance on PASCAL VOC 2012 dataset. Code is available at https://github.com/Wu0409/HSC_WSSS.

Service placement constitutes a prominent subject in edge computing, given its pivotal role in mitigating service latency. However, conventional heuristic placement strategies encounter difficulties in coping with the intricate dynamics of constantly evolving real-world settings, attributable to the non-stationary nature of service demand and time-varying network conditions. To tackle this challenge, deep reinforcement learning (DRL) techniques have received substantial attention. These approaches harness dynamic edge environments, encompassing cloud, edge, terminal devices, and wireless communication channels, to acquire placement policies through interactive learning. Nevertheless, a significant drawback of these learning-based approaches lie in their frequent necessity for retraining when confronted with novel edge environments, thereby resulting in inefficiencies in dynamic service placement across heterogeneous environments. To overcome this limitation, this paper proposes an innovative recurrent neural network (RNN)-based meta-reinforcement learning (meta-RL) technique. This technique capitalizes on prior historical experiences to expedite the learning of new policies in unfamiliar situations. Extensive simulation results are presented to illustrate the commendable performance of our proposed scheme.

Federated learning (FL) has been widely used in edge computing that enables artificial intelligence at the network edge as a distributed machine learning paradigm. In contrast to traditional cloud-based distributed training, the heterogeneity in edge computing may cause federated learning taking long training time. In this paper, we adapt control parameter (i.e., local epoch size) across devices to minimize wall-clock convergence time with joint consideration of resource heterogeneity and statistical heterogeneity. To analyze the influence of statistical heterogeneity, we derive a convergence upper bound for synchronous FL algorithm and establish the relationship between the number of training rounds and local epoch size under heterogeneous data distribution. Based on the convergence bound, we can solve the non-convex problem of minimizing FL training time with accuracy constraint and obtain near-optimal local epoch size. We develop a scheduling algorithm that estimates the statistical heterogeneity in initial training rounds and subsequently guides adaptive local training across devices. Practically, we evaluate our algorithm in a variety of heterogeneous scenarios. Extensive simulation results demonstrate that our algorithm performs high convergence speed over wall-clock time and spends less time reaching target accuracy compared with benchmark approaches.

Robust loss functions are essential for training deep neural networks in the presence of noisy labels. Some robust loss functions use Mean Absolute Error (MAE) as its necessary component. For example, the recently proposed Active Passive Loss (APL) uses MAE as its passive loss function. However, MAE treats every sample equally, slows down the convergence and can make training difficult. In this work, we propose a new class of theoretically robust passive loss functions different from MAE, namely *Normalized Negative Loss Functions* (NNLFs), which focus more on memorized clean samples. By replacing the MAE in APL with our proposed NNLFs, we improve APL and propose a new framework called *Active Negative Loss* (ANL). Experimental results on benchmark and real-world datasets demonstrate that the new set of loss functions created by our ANL framework can outperform state-of-the-art methods. The code is available at https://github.com/Virusdoll/Active-Negative-Loss.

Multi-access Edge Computing has come forth as a promising paradigm to provide low-latency computing service to mobile end users. Its basic idea is to deploy computation resources at the edge of core networks such as wireless access points, and then users can offload their tasks to nearby edge nodes for processing. Plenty of works have well studied the task offloading problem, aiming to reduce task completion delays. Also, a few recent works have focused on task dispatching among edge nodes to balance their workloads and improve resource utilization. In this work, we jointly consider the task offloading and dispatching problem in an edge computing system with interconnected access points. Furthermore, we assume both end devices and access points are rational, which only care about their own benefits. To solve the joint problem, we firstly formulate it as a multi-leader multi-follower Stackelberg game, and rigorously prove the existence of a Stackelberg equilibrium. Then, we propose two algorithms for task offloading and dispatching, respectively. Extensive simulations are conducted to show the superiority of our proposed approach. We also demonstrate that an upper bound with a constant approximation ratio is achieved by our approach.

AI-empowered 5G/6G networks play a substantial role in taking full advantage of the Internet of Things (IoT) to perform complex computing by offloading tasks to edge services deployed in intelligent transport systems. However, offloading behavior has a certain regularity, and the real-time location of users can easily be inferred by attackers who have historical user data during the data transmission process. To address this problem, a privacy-oriented task offloading method that can resist attacks from privacy attackers with prior knowledge is proposed. First, the local computing model, channel model, and privacy loss model are defined and used to quantify evaluation indicators, such those related to privacy, time, and energy. Among them, privacy loss is formalized as the probability of a successful attack by an attacker with prior knowledge. Second, the process of solving an optimal task offloading decision problem is formalized into a Markov decision process (MDP). Finally, the deep reinforcement learning (DRL) method PPO2 is proposed to solve the planning problem of task offloading with good generalization and convergence speed, where we focus on the location privacy requirement. Experiments show that our method can handle large-scale task offloading and obtain offloading policies with reduced privacy loss, energy consumption and time delays.

Due to the urgent emergence of computation-intensive intelligent applications on end devices, edge computing has been put forward as an extension of cloud computing, to satisfy the low-latency requirements of these applications. To process heterogenous computation tasks on an edge node, the corresponding services should be placed in advance, including installing softwares and caching databases/libraries. Considering the limited storage space and computation resources on the edge node, services should be elaborately selected and deployed on the edge node and its computation resources should be carefully allocated to placed services, according to the arrivals of computation workloads. The joint service placement and computation resource allocation problem is particularly complicated, in terms of considering the stochastic arrivals of tasks, the additional latency incurred by service migration, and the waiting time of unprocessed tasks. Benefiting from deep reinforcement learning, we propose a novel approach based on parameterized deep Q networks to make the joint service placement and computation resource allocation decisions, with the objective of minimizing the total latency of tasks in a long term. Extensive simulations are conducted to evaluate the convergence and performance achieved by our proposed approach.

With the development of 5G, edge computing has raised as a promising technology to satisfy the requirements of computation-intensive and delay-sensitive applications. In this work, we try to propose a task offloading strategy for end users, with considering their individual rationality and collective rationality at the same time. Specially, a user only with individual rationality aims to minimize the completion time of its own task, while a user only with collective rationality aims to minimize the total task completion time achieved by the system. The problem is particularly difficult, as there exist essential conflicts between the individual utility of each user and the collective utility of the system, which cannot be optimized simultaneously. To overcome the difficulties, we first reformulate the problem as a coalition formation game. Then, we propose an iterative algorithm, in which each user can make its task offloading decision in a decentralized way. Additionally, we rigorously prove the properties achieved by our algorithm in terms of stability, optimality, and convergence rate. Extensive simulations are also conducted to validate the performance of our algorithm compared with baselines.

Federated Learning (FL) is a distributed machine learning technology that trains models on large-scale distributed devices while keeping training data localized and privatized. However, in settings where data is distributed in a not independent and identically distributed (non-I.I.D.) fashion, the single joint model produced by FL suffers in terms of test set accuracy and communication costs. And a multi-layer topology are widely deployed for FL in real scenarios. Therefore, we propose FedBox, a chunked local aggregation federated learning framework to improve the generalization ability and aggregation efficiency of model in non-I.I.D. data by adapting to the topology of the real network. Moreover, we study the adaptive gradient descent (AGC) to mitigate the feature shift caused by training non-I.I.D. data. In this work, we modified the aggregation strategy of FL by introducing a virtual node layer based on local stochastic gradient methods (SGD), and separate the edge node cluster by the similarity between the local update model and the global update model. We show that FedBox can effectively improve convergence speed and test accuracy, while reducing communication cost. Training results on FederatedEMNIST, Cifar10, Cifar100 and Shakespeare datasets indicate that FedBox allows model training to converge in fewer communication rounds and improves training accuracy by up to 3.1% compared with FedAVG. In addition, we make an empirical analysis of the extended range of virtual nodes.

The task of next Point-of-Interest (POI) recommendation aims at recommending a list of POIs for a user to visit at the next timestamp based on his/her previous interactions, which is valuable for both location-based service providers and users. Recent state-of-the-art studies mainly employ recurrent neural network (RNN) based methods to model user check-in behaviors according to user’s historical check-in sequences. However, most of the existing RNN-based methods merely capture geographical influences depending on physical distance or successive relation among POIs. They are insufficient to capture the high-order complex geographical influences among POI networks, which are essential for estimating user preferences. To address this limitation, we propose a novel Graph-based Spatial Dependency modeling (GSD) module, which focuses on explicitly modeling complex geographical influences by leveraging graph embedding. GSD captures two types of geographical influences, i.e., distance-based and transition-based influences from designed POI semantic graphs. Additionally, we propose a novel Graph-enhanced Spatial-Temporal network (GSTN), which incorporates user spatial and temporal dependencies for next POI recommendation. Specifically, GSTN consists of a Long Short-Term Memory (LSTM) network for user-specific temporal dependencies modeling and GSD for user spatial dependencies learning. Finally, we evaluate the proposed model using three real-world datasets. Extensive experiments demonstrate the effectiveness of GSD in capturing various geographical influences and the improvement of GSTN over state-of-the-art methods.

Due to the explosion of mobile devices and the evolution of wireless communication technologies, novel applications with intensive computation demands and low-latency requirements have arisen. Edge computing has been proposed as an extension of cloud computing, which moves computation workloads from remote cloud to network edge. Cooperating edge computing and cloud computing can significantly reduce the latency of computation tasks. However, considering the heterogeneity and stochastic arrivals of tasks and the limited computation and communication resources on the edge, task offloading and resource allocation are two joint crucial problems in an edge-cloud orchestrated computing system. In this paper, we propose an online task offloading and resource allocation approach for edge-cloud orchestrated computing, with the aim to minimize the average latency of tasks over time. We first build system models to analyze the latency and energy consumption incurred under different computing modes and formally formulate the joint problem as a mixed-integer optimal decision problem. Then, we employ Lyapunov optimization and duality theory to decompose the problem into a set of subproblems, which can be solved in a semi-decentralized way. We also formally analyze that our approach can achieve near-optimal performance. Extensive simulations are conducted to verify the superiority of our approach.

The emerging IoT stream processing is a key enabling technology for the time-critical IoT applications, which often require high accuracy and low latency. Existing stream processing engines are insufficient to meet these requirements, since they could not integrate and respond timely to variable network conditions in the dynamic wireless environment. Recent efforts focusing on adaptive streaming support user-specified policies to adapt to the variable network conditions. However, those manual-policies can hardly achieve optimal performance across a broad set of network conditions and quality of experience (QoE) objectives. In this paper, we present a Reinforcement Learning-based Adaptive streaming system (RL-Adapt) that is capable of generating adaption policies using RL-strategy and providing declarative APIs for efficient development. RL-Adapt trains a neural network model that can automatically select the optimal policy based on the observed network conditions. RL-Adapt does not rely on pre-defined models or assumptions on the environment. Instead, it learns to make decisions solely through observations of the resulting performance of past decisions. We implemented RL-Adapt and evaluated its performance extensively in three representative real-world IoT applications. Our results show that RL-Adapt outperforms the state-of-the-art scheme, with 20% improvements on average QoE.

With the explosion of mobile smart devices, many computation intensive applications have emerged, such as interactive gaming and augmented reality. Mobile-edge computing (EC) is put forward, as an extension of cloud computing, to meet the low-latency requirements of the applications. In this article, we consider an EC system built in an ultradense network with numerous base stations. Heterogeneous computation tasks are successively generated on a smart device moving in the network. An optimal task offloading strategy, as well as optimal CPU frequency and transmit power scheduling, is desired by the device user to minimize both task completion latency and energy consumption in a long term. However, due to the stochastic task generation and dynamic network conditions, the problem is particularly difficult to solve. Inspired by reinforcement learning, we transform the problem into a Markov decision process. Then, we propose an attention-based double deep Q network (DDQN) approach, in which two neural networks are employed to estimate the cumulative latency and energy rewards achieved by each action. Moreover, a context-aware attention mechanism is designed to adaptively assign different weights to the values of each action. We also conduct extensive simulations to compare the performance of our proposed approach with several heuristic and DDQN-based baselines.

Multi-access Edge Computing (MEC) comes forth as a promising computing paradigm to meet the low-delay requirements of computation-intensive applications. In this work, we focus on the task offloading and dispatching problem in MEC, where mobile devices (MDs) can decide to execute tasks locally or offload them to access points (APs), and each AP can dispatch its received tasks to other APs. Specially, we consider both APs and MDs are selfish, which aim to minimize their respective task completion delay. The problem is particularly difficult, as there exists computing resource competition among MDs and APs, respectively. Furthermore, the offloading decisions made by MDs and the dispatching decisions made by APs are interactive. To overcome the challenges, we first formulate the problem as a multi-leader multi-follower Stackelberg game, and rigorously prove the existence of a Stackelberg equilibrium. Then, we propose an efficient approach to achieve a Stackelberg equilibrium, which includes a Q-learning based offloading strategy for MDs and a best response based dispatching strategy for APs. We also demonstrate an upper bound of the total completion delay achieved by our approach with a constant approximation ratio. Extensive simulations are also conducted to show the performance of our approach, compared with baselines.

With the explosion of mobile smart devices, many computation intensive applications have emerged, such as interactive gaming and augmented reality. Mobile edge computing is put forward, as an extension of cloud computing, to meet the low-latency requirements of the applications. In this paper, we consider an edge computing system built in an ultra-dense network with numerous base stations, and heterogeneous computation tasks are successively generated on a smart device moving in the network. An optimal task offloading strategy, as well as optimal CPU frequency and transmit power scheduling, is desired by the device user, to minimize both task completion latency and energy consumption in a long-term. However, due to the stochastic computation tasks and dynamic network conditions, the problem is particularly difficult to solve. Inspired by reinforcement learning, we transform the problem into a Markov decision process. Then, we propose an online offloading approach based on a double deep Q network, in which a specific neural network model is also provided to estimate the cumulative reward achieved by each action. We also conduct extensive simulations to compare the performance of our proposed approach with baselines.

As an environment-friendly public transport, shared bikes have become an important urban transport, providing cheap and convenient services for urban residents. However, the number of docks of a station in a bike sharing system is fixed when it is built, and there exists imbalance between bike usage and supply in reality. An accurate real-time free dock prediction can help guide users to choose a proper station (with free bikes/docks) to rent or return a bike. Many earlier efforts are paid to do bike sharing prediction based on model-based approaches. Recently, deep neural networks (DNN), like convolutional neural networks (CNN) and recurrent neural networks (RNN), have been introduced to solve traffic prediction problems. However, three are some unsolved issues to make accurate real-time free dock prediction, such as learning complicated temporal variation and periodicity of bike usage, spatial correlations of free docks among different stations, and the impact of external factors like weather. To overcome these challenges, we propose a novel deep neural network model, which combines graph convolution and a residual structure together. We first model a bike sharing system as a weighted graph, and the non-Euclidean spatial correlations among stations (represented by weighted edges in the graph) are extracted by random walk operation in graph convolution layers. Moreover, periodic patterns of free docks in different time scales are captured by a residual structure, and external factors are considered to improve the accuracy of prediction. We also conduct comprehensive experiments based on a public real-world dataset of riding trips from Boston, and the results show that our method outperforms state-of-the-art baselines.

Mobile edge computing (MEC) is a promising paradigm to meet the low latency requirement of applications like virtual/augmented reality, which moves computation workloads from remote cloud to network edge. Task offloading is a crucial problem in a MEC system, which significantly concerns with the achieved latency of tasks and the energy consumption on edge nodes. Although many efforts have been paid on the problem, most of them ignore the hierarchical architecture of MEC and the heterogeneity of tasks generated by different applications. In this paper, we consider a three-layer MEC system with tasks of various types stochastically arrive in real time, in which each task can be executed on mobile smart devices, offloaded to the edge server, or offloaded to the remote cloud. The task offloading problem in such a MEC system with the objective to minimize the averaged latency over time is NP-hard. To solve it, we employ Lyapunov optimization and duality theory to reformulate the problem and decompose it into a set of subproblems. Each subproblem can be distributedly solved by a mobile device or the edge server individually. Extensive simulations are also performed to verify the feasibility and superiority of our approach.

Instability and slowness are two main problems in deep reinforcement learning. Even if proximal policy optimization (PPO) is the state of the art, it still suffers from these two problems. We introduce an improved algorithm based on proximal policy optimization, mixed distributed proximal policy optimization (MDPPO), and show that it can accelerate and stabilize the training process. In our algorithm, multiple different policies train simultaneously and each of them controls several identical agents that interact with environments. Actions are sampled by each policy separately as usual, but the trajectories for the training process are collected from all agents, instead of only one policy. We find that if we choose some auxiliary trajectories elaborately to train policies, the algorithm will be more stable and quicker to converge especially in the environments with sparse rewards.

As metropolises develop, air pollution has become a serious problem, especially in developing countries like China. Many governments and researchers have devoted themselves to tackling and solving this problem. With the proliferation of smartphones, mobile crowdsensing is becoming a promising paradigm for monitoring large-scale environmental phenomena. In a practical crowdsensing system, incentives should be provided to encourage the participation of rational smartphone users, because it incurs various costs on users to collect sensing data. However, monitoring fine-grained air pollution in a large urban area based on crowdsensing will lead to high payments, which makes designing an efficient incentive mechanism a challenging problem. Fortunately, compressive sensing (CS) has been proved as an effective technology to reduce the amount of collected data via exploiting the spatial correlations among sensing data. In this article, we employ CS in the air pollution monitoring application, in which only a sampled set of locations are selected to collect data and provide incentives to the participants, and air pollution concentrations in unselected locations are inferred via CS. We propose an active learning scheme, which iteratively selects valuable locations to collect sensing data. Moreover, an expectation maximization-based algorithm is designed to detect the contexts in which sensing data are collected, and an efficient incentive mechanism is provided to encourage users with low costs participating. Comprehensive simulations are conducted to demonstrate the performance of our proposed scheme.

It is of significant importance to provide incentives to smartphone users in mobile crowd sensing systems. Recently, a number of auction-based incentive mechanisms have been proposed. However, an auction-based incentive mechanism may unexpectedly release the location privacy of smartphone users, which may seriously reduce the willingness of users participating in contributing sensing data. In an auction-based incentive mechanism, even if the location of a user is not enclosed in his/her bid submitted to the platform, the location information may still be inferred by an adversary by using the prices of the tasks required by the user. We take an example to show how an attack can recover the location information of a smartphone user by merely knowing his/her bid. To defend against such an attack, we propose a method to protect location privacy in auctions for mobile crowd sensing systems. This method encrypts prices in a bid so that the adversary cannot access and hence the location privacy of users can be protected. In the meanwhile, however, the auction can proceed properly, i.e. the platform can select the user offering the lowest price for each sensing task or the platform can choose users with budget constraint. We demonstrate the effectiveness of our proposed method with theoretical analysis and simulations.

We consider the crucial problem of maximizing the social welfare of a vehicular participatory sensing system, where the system's social welfare is measured by the amount of sensing data delivered to a central platform through a vehicular ad hoc network. The key to the problem is to control network stability since both network congestion and idleness will slump system social welfare. However, several great challenges exist. First, limited vehicle-to-vehicle (V2V) link capacity and vehicle buffer size will lead to heavy network congestion when each individual vehicle blindly injects too much data into the network hoping to get more rewards. Second, the highly dynamic network topology and stochastic inter-vehicle contacts have a serious impact on the performance of multi-hop data transmission. Third, vehicles need to be practically rewarded based on their sensing and transmission cost, which, however, greatly vary among vehicles. To tackle the aforementioned challenges, we propose a distributed backpressure control approach, the first work to the best of our knowledge, to maximize the social welfare while balancing network stability for a vehicular participatory sensing system. Combining vehicular network properties and Lyapunov optimization techniques, individualized strategies are developed for each participant to control its sensing rate, make its own routing decisions, and set its own price for data relaying. Formally proved by rigorous theoretical analysis, the social welfare achieved by the proposed approach is comparative to the optimum performance. In addition, extensive data-driven simulations based on real taxi GPS traces have been conducted, and the results confirm the efficacy of the proposed algorithm.

With the development of distributed energy generators and storages, the sustainability of a microgrid comprised of multiple electricity users is significantly increased. Maximizing the efficiency of generated renewable energy is vital to running a sustainable microgrid as it indicates reducing the usage of thermal electricity purchased from the macrogrid. To this end, the excessive renewable energy of a user should be shared with others who are short of energy. Unfortunately, coordinating the transfers of renewable energy among the users in a microgrid is particularly difficult, given the rational nature of users, the stochastic nature of renewable energy and the dynamic nature of energy demand of each user. In this paper, we consider the coupled problem of maximizing the renewable energy efficiency of a sustainable microgrid as well as stimulating rational users to share excessive renewable energy. We propose a near-optimal scheduling algorithm, which determines the amounts of renewable energy transferred among users in an online fashion. We also design an efficient pricing mechanism for the trade of energy among users based on double auction. We rigorously prove that our online scheduling algorithm is approximately optimal and the pricing mechanism guarantees the property of individual rationality of users. Comprehensive simulation results demonstrate the efficacy of our online algorithm and incentive mechanism.

In this paper, we investigate the problem of energy balanced data collection in wireless sensor networks, aiming to balance energy consumption among all sensor nodes during the data propagation process. Energy balanced data collection can potentially save energy consumption and prolong network lifetime, and hence, it has many practical implications for sensor network design and deployment. The traditional hop-by-hop transmission model allows a sensor node to propagate its packets in a hop-by-hop manner toward the sink, resulting in poor energy balancing for the entire network. To address the problem, we apply a slice-based energy model, and divide the problem into inter-slice and intra-slice energy balancing problems. We then propose a probability-based strategy named inter-slice mixed transmission protocol and an intra-slice forwarding technique to address each of the problems. We propose an energy-balanced transmission protocol by combining both techniques to achieve total energy balancing. In addition, we study the condition of switching between inter-slice transmission and intra-slice transmission, and the limitation of hops in an intra-slice transmission. Through our extensive simulation studies, we demonstrate that the proposed protocols achieve energy balancing, prolong network lifespan, and decrease network delay, compared with the hop-by-hop transmission and a cluster-based routing protocol under various parameter settings.

As air pollution is becoming a serious problem in developing nations, governments try to track and solve this problem by monitoring air pollution. With the proliferation of smartphones, mobile crowdsensing becomes a promising paradigm for monitoring fine-grained air pollution in urban areas. As existing studies have shown that pollutant concentrations have inherent spatiotemporal correlations, compressive sensing is an effective technology to reduce the amount of data collected through crowdsensing. In a practical crowdsensing application, incentives are expected by smartphone users for contributing sensing data. However, how to design incentives to collect high- quality sensing data with low costs is difficult in compressive crowdsensing. In this work, we propose an iterative scheme for the process of crowdsensing-based air pollution monitoring, where incentives are updated online according to the distribution of collected sensing data. Comprehensive simulations have been conducted to demonstrate the efficacy of our proposed scheme.

It is of significant importance to provide incentives to smartphone users in mobile crowd sensing systems. And a number of auction-based incentive mechanisms have been proposed. However, an auction-based incentive mechanism may unexpectedly release the location privacy of smartphone users, which may seriously reduce users' willingness of participating in mobile crowd sensing. In an auction-based mechanism, even if the location of the user is not enclosed in its bid submitted to the platform, the location information may still be inferred by an adversary by using the prices of the tasks required by the user. We show such an attack on a typical auction-based incentive mechanism and reveal that the attack can recover the location information of a smartphone user by merely knowing the bid from the user. To defend against such an attack, we propose P2, a location privacy-preserving auction mechanism for mobile crowd sensing systems. This mechanism encrypts prices in a bid so that the adversary cannot access and hence the location privacy of the user can be protected. In the meanwhile, however, the auction can proceed properly, i.e. the platform can select the user offering the lowest price for each sensing task. We demonstrate the effectiveness of our the proposed mechanism with simulation experiments.

With the development of renewable energy generators and electricity storages, microgrids become a promising technology of the smart grid. Maximizing the usage of renewable energy is vital to running a microgrid as it indicates reduction of the usage of thermal electricity purchased from the macrogrid. To this end, the excessive renewable energy of a user should be transferred to other users who need energy. Unfortunately, coordinating the transfers of renewable energy among the users in the microgrid is particularly difficult due to the stochastic nature of renewable energy, and the dynamic energy demand of each user. In this paper, we consider the problem of maximizing the long-term renewable energy usage by exchanging excessive renewable energy among users in a microgrid. We propose an online control algorithm which determines the amounts of renewable energy transferred among users in an online fashion. We rigorously prove that our online control algorithm is approximately optimal. We have conducted comprehensive simulation results that demonstrate the efficacy of our online algorithm.

We consider the crucial problem of maximizing the system-wide performance which takes into account request processing throughput, smartphone user experience and system stability in a participatory sensing system with cooperative smartphones. Three important controls need to be made, i.e., 1) request admission control, 2) task allocation, and 3) task scheduling on smartphones. It is highly challenging to achieve the optimal system-wide performance, given arbitrary unknown arrivals of sensing requests, intrinsic tradeoff between request processing throughput and smartphone user experience degradation, and heterogenous requests. Little existing work has studied this crucial problem of maximizing the system-wide performance of a participatory sensing system as a whole. In response to the challenges, we propose an optimal online control approach to maximize the system-wide performance of a participatory sensing system. Exploiting the stochastic Lyapunov optimization techniques, it derives the optimal online control strategies for request admission control, task allocation and task scheduling on smartphones. The most salient feature of our approach is that the achieved system-wide performance is arbitrarily close to the optimum, despite unpredictable and arbitrary request arrivals. Rigorous theoretical analysis and comprehensive simulation evaluation jointly demonstrate the efficacy of our online control approach.

Many cities suffer from noise pollution, which compromises people's working efficiency and even mental health. New York City (NYC) has opened a platform, entitled 311, to allow people to complain about the city's issues by using a mobile app or making a phone call; noise is the third largest category of complaints in the 311 data. As each complaint about noises is associated with a location, a time stamp, and a fine-grained noise category, such as "Loud Music" or "Construction", the data is actually a result of "human as a sensor" and "crowd sensing", containing rich human intelligence that can help diagnose urban noises. In this paper we infer the fine-grained noise situation (consisting of a noise pollution indicator and the composition of noises) of different times of day for each region of NYC, by using the 311 complaint data together with social media, road network data, and Points of Interests (POIs). We model the noise situation of NYC with a three dimension tensor, where the three dimensions stand for regions, noise categories, and time slots, respectively. Supplementing the missing entries of the tensor through a context-aware tensor decomposition approach, we recover the noise situation throughout NYC. The information can inform people and officials' decision making. We evaluate our method with four real datasets, verifying the advantages of our method beyond four baselines, such as the interpolation-based approach.

In this paper we study a social welfare maximization problem for spectrum sharing in cognitive radio networks. To fully use the spectrum resource, the spectrum owned by the licensed primary user (PU) can be leased to secondary users (SUs) for transmitting data. We first formulate the social welfare of a cognitive radio network, considering the cost for the primary user sharing spectrum and the utility gained for secondary users transmitting data. The social welfare maximization is a convex optimization, which can be solved by standard methods in a centralized manner. However, the utility function of each secondary user always contains the private information, which leads to the centralized methods disabled. To overcome this challenge, we propose an iterative distributed algorithm based on a pricing-based decomposition framework. It is theoretically proved that our proposed algorithm converges to the optimal solution. Numerical simulation results are presented to show that our proposed algorithm achieves optimal social welfare and fast convergence speed.

There are increasing interests on mobile opportunistic networks which have promising applications. Constructing a mobile backbone can effectively improve the packet delivery performance of a mobile opportunistic network by excluding poor relay nodes and reducing packet collisions. However, it is highly challenging to construct an effective mobile backbone because of the absence of the quantitative relationship between the network performance and the selection of backbone nodes, and expositive search space. As nodes exhibit clear sociality observed in previous studies, We explicitly take such node sociality into account when computing the backbone for mobile opportunistic networks and we incrementally propose three algorithms for computing the mobile backbone. Trace-driven simulations have been conducted and simulation results demonstrate that the sociality-aware algorithms can achieve low delivery delay and high delivery ratio.