In federated learning (FL), distributed users collaboratively train a neural network model under the coordination of a central server. However, time-varying client availability, coupled with non-independent and non-identically distributed (non-IID) datasets, leads to a biased convergence. In this work, we prove the convergence of FL under time-varying client availability. The theoretical result shows that biased convergence occurs when available client distribution does not algin with the client population distribution. To address this challenge, we propose a pricing-based incentive mechanism to encourage clients to adjust their availability. First, we model the strategic interactions among clients as a non-cooperative game under an arbitrary pricing scheme. We prove that this game is a potential game and its equilibrium can be found through optimization. Second, we derive an optimal pricing scheme for large client populations and propose a bi-level optimization algorithm using Particle Swarm Optimization (PSO) for general scenarios. Through analysis of client availability evolution, we prove the effectiveness of our scheme in mitigating biased convergence. Experimental results using real-world client availability dataset show that our approach addresses time-varying client availability issue, achieving up to 99.5% improvement over benchmarks and enhancing FL convergence rates by up to 2.49 times.

Fairness in federated learning (FL) has emerged as a critical concern, aiming to develop an unbiased model among groups (e.g., male or female) of diverse sensitive features. However, there is a trade-off between model performance and fairness, i.e., improving fairness will decrease performance. Existing approaches have characterized such a trade-off by introducing hyperparameters to quantify client’s preferences over fairness and performance. Nevertheless, these approaches are limited to scenarios where each client has only a single pre-defined preference, and fail to address in cases where each client has multiple preferences. In this work, we aim to design algorithms that allow the trained model to adapt to each client’s diverse preferences in real time. The key challenges lie in (I) associating preferences with the trained model; (II) mitigating data heterogeneity; (III) preventing preference leakage; and (IV) handling data noise. To address these, we propose two preference-aware schemes, PraFFL and PraFFL-R, designed to generate models tailored to specific client preferences. PraFFL tackles challenges (I)–(III) by incorporating a hypernetwork that adaptively adjusts the model according to each client’s preferences, thereby better satisfying individual needs. To further address data noise (challenge (IV)), we introduce PraFFL-R, which enhances the robustness of the learned Pareto front by optimizing worst-case scenarios among preference-specific models. We provide theoretical guarantees demonstrating that PraFFL and PraFFL-R can produce an optimal model customized for any client preference, with proofs of linear convergence in the strongly convex setting and sublinear convergence in the non-convex setting. Experimental results show that our proposed PraFFL and PraFFL-R outperform five fair FL algorithms in terms of the model’s capability of adapting to clients’ different preferences.

AI-Generated Content (AIGC) based on Generative Diffusion Models (GDMs) is revolutionizing content creation and promoting substantial advancements in domains like autonomous driving and robotics. Leveraging progress in Mobile Edge Computing (MEC) and model compression techniques, GDMs are increasingly being deployed on Edge Servers (ESs) and User Equipments (UEs), which typically face resource limitations. In such MEC-enabled scenarios, designing an efficient inference scheme for GDMs still remains a significant challenge, due to the resource constraints on ESs and UEs as well as the personalized demands of AIGC users. In this work, we propose a novel hybrid inference scheme, which consists of two stages: public prompt generation and common-to-personalized inference. In the first stage, a Large Language Model (LLM) is adopted to generate public prompts derived from the common features of users' personal prompts. In the second stage, a common inference phase based on public prompts is first executed for all users (to produce common intermediate results), and then a personalized inference phase based on each user's personal prompts is performed for each individual user (to generate final contents). Clearly, by introducing the common inference phase, the total inference steps can be significantly reduced. In such a scheme, we further study a hybrid inference optimization problem to optimize both common and personalized inference steps, aiming to maximize the total Quality of Service (QoS), while minimizing delay and energy consumption. Simulation results show that our proposed scheme significantly outperforms existing benchmarks, with the performance gains ranging from 12.6 % to 102.2 %.

Machine unlearning strives to uphold the data owners' right to be forgotten by enabling models to selectively forget specific data. Recent advances suggest pre-computing and storing statistics extracted from second-order information and implementing unlearning through Newton-style updates. However, the Hessian matrix operations are extremely costly and previous works conduct unlearning for empirical risk minimizer with the convexity assumption, precluding their applicability to high-dimensional over-parameterized models and the nonconvergence condition. In this paper, we propose an efficient Hessian-free unlearning approach. The key idea is to maintain a statistical vector for each training data, computed through affine stochastic recursion of the difference between the retrained and learned models. We prove that our proposed method outperforms the state-of-the-art methods in terms of the unlearning and generalization guarantees, the deletion capacity, and the time/storage complexity, under the same regularity conditions. Through the strategy of recollecting statistics for removing data, we develop an online unlearning algorithm that achieves near-instantaneous data removal, as it requires only vector addition. Experiments demonstrate that our proposed scheme surpasses existing results by orders of magnitude in terms of time/storage costs with millisecond-level unlearning execution, while also enhancing test accuracy.

Learning-based street scene semantic understanding in autonomous driving (AD) has advanced significantly recently, but the performance of the AD model is heavily dependent on the quantity and quality of the annotated training data. However, traditional manual labeling involves high cost to annotate the vast amount of required data for training robust model. To mitigate this cost of manual labeling, we propose a Label Anything Model (denoted as LAM), serving as an interpretable, high-fidelity, and prompt-free data annotator. Specifically, we firstly incorporate a pretrained Vision Transformer (ViT) to extract the latent features. On top of ViT, we propose a semantic class adapter (SCA) and an optimization-oriented unrolling algorithm (OptOU), both with a quite small number of trainable parameters. SCA is proposed to fuse ViT-extracted features to consolidate the basis of the subsequent automatic annotation. OptOU consists of multiple cascading layers and each layer contains an optimization formulation to align its output with the ground truth as closely as possible, though which OptOU acts as being interpretable rather than learning-based blackbox nature. In addition, training SCA and OptOU requires only a single pre-annotated RGB seed image, owing to their small volume of learnable parameters. Extensive experiments clearly demonstrate that the proposed LAM can generate high-fidelity annotations (almost 100% in mIoU) for multiple real-world datasets (i.e., Camvid, Cityscapes, and Apolloscapes) and CARLA simulation dataset.

Recent methods leverage a hypernet to handle the performance-fairness trade-offs in federated learning. This hypernet maps the clients' preferences between model performance and fairness to preference-specifc models on the trade-off curve, known as local Pareto front. However, existing methods typically adopt a uniform preference sampling distribution to train the hypernet across clients, neglecting the inherent heterogeneity of their local Pareto fronts. Meanwhile, from the perspective of generalization, they do not consider the gap between local and global Pareto fronts on the global dataset. To address these limitations, we propose HetPFL to effectively learn both local and global Pareto fronts. HetPFL comprises Preference Sampling Adaptation (PSA) and Preference-aware Hypernet Fusion (PHF). PSA adaptively determines the optimal preference sampling distribution for each client to accommodate heterogeneous local Pareto fronts. While PHF performs preference-aware fusion of clients' hypernets to ensure the performance of the global Pareto front. We prove that HetPFL converges linearly with respect to the number of rounds, under weaker assumptions than existing methods. Extensive experiments on four datasets show that HetPFL significantly outperforms seven baselines in terms of the quality of learned local and global Pareto fronts.

To improve the generalization of the autonomous driving (AD) perception model, vehicles need to update the model over time based on the continuously collected data. As time progresses, the amount of data fitted by the AD model expands, which helps to improve the AD model generalization substantially. However, such ever-expanding data is a double-edged sword for the AD model. Specifically, as the fitted data volume grows to exceed the AD model’s fitting capacities, the AD model is prone to under-fitting. To address this issue, we propose to use a pretrained Large Vision Models (LVMs) as backbone coupled with downstream perception head to understand AD semantic information. This design can not only surmount the aforementioned under-fitting problem due to LVMs’ powerful fitting capabilities, but also enhance the perception generalization thanks to LVMs’ vast and diverse training data. On the other hand, to mitigate vehicles’ computational burden of training the perception head while running LVM backbone, we introduce a Posterior Optimization Trajectory (POT)-Guided optimization scheme (POTGui) to accelerate the convergence. Concretely, we propose a POT Generator (POTGen) to generate posterior (future) optimization direction in advance to guide the current optimization iteration, through which the model can generally converge within 10 epochs. Extensive experiments demonstrate that the proposed method improves the performance by over 66.48% and converges faster over 6 times, compared to the existing state-of-the-art approaches.

Street Scene Semantic Understanding (denoted as S3U) is a crucial but complex task for autonomous driving (AD) vehicles. Their inference models typically face poor generalization due to domain-shift. Federated Learning (FL) has emerged as a promising paradigm for enhancing the generalization of AD models through privacy-preserving distributed learning. However, these FL AD models face significant temporal catastrophic forgetting when deployed in dynamically evolving environments, where continuous adaptation causes abrupt erosion of historical knowledge. This paper proposes Federated Exponential Moving Average (FedEMA), a novel framework that addresses this challenge through two integral innovations: (I) Server-side model’s historical fitting capability preservation via fusing current FL round’s aggregation model and a proposed previous FL round’s exponential moving average (EMA) model; (II) Vehicle-side negative entropy regularization to prevent FL models’ possible overfitting to EMA-introduced temporal patterns. Above two strategies empower FedEMA a dual-objective optimization that balances model generalization and adaptability. In addition, we conduct theoretical convergence analysis for the proposed FedEMA. Extensive experiments both on Cityscapes dataset and Camvid dataset demonstrate FedEMA’s superiority over existing approaches, showing 7.12% higher mean Intersectionover-Union (mIoU).

Fairness in federated learning has emerged as a critical concern, aiming to develop an unbiased model among groups (e.g., male or female) of diverse sensitive features. However, there is a trade-off between model performance and fairness, i.e., improving model fairness will decrease model performance. Existing approaches have characterized such a trade-off by introducing hyperparameters to quantify client's preferences for model fairness and model performance. Nevertheless, these approaches are limited to scenarios where each client has only a single pre-defined preference, and fail to work in practical systems where each client generally has multiple preferences. To this end, we propose a Preference-aware scheme in Fair Federated Learning (called PraFFL) to generate preference-specific models in real time. PraFFL can adaptively adjust the model based on each client's preferences to meet their needs. We theoretically prove that PraFFL can offer the optimal model tailored to an arbitrary preference of each client, and show its linear convergence. Experimental results show that our proposed PraFFL outperforms six fair federated learning algorithms in terms of the model's capability of adapting to clients' different preferences. Our implementation is available at https://github.com/rG223/PraFFL.

Chain-of-Thought (CoT) prompting improves the reasoning performance of large language models (LLMs) by encouraging step-by-step thinking. However, CoT-based methods depend on intermediate reasoning steps, which limits scalability and generalization. Recent work explores recursive reasoning, where LLMs reuse internal layers across iterations to refine latent representations without explicit CoT supervision. While promising, these approaches often require costly pretraining and lack a principled framework for how reasoning should evolve across iterations. We address this gap by introducing **Flow Chain-of-Thought (Flow CoT)**, a reasoning paradigm that models recursive inference as a progressive trajectory of latent cognitive states. Flow CoT frames each iteration as a distinct cognitive stage—deepening reasoning across iterations without relying on manual supervision. To realize this, we propose **SCOUT** (*Stepwise Cognitive Optimization Using Teachers*), a lightweight fine-tuning framework that enables Flow CoT-style reasoning without the need for pretraining. SCOUT uses progressive distillation to align each iteration with a teacher of appropriate capacity, and a cross-attention-based retrospective module that integrates outputs from previous iterations while preserving the model’s original computation flow. Experiments across eight reasoning benchmarks show that SCOUT consistently improves both accuracy and explanation quality, achieving up to 1.8\% gains under fine-tuning. Qualitative analyses further reveal that SCOUT enables progressively deeper reasoning across iterations—refining both belief formation and explanation granularity. These results not only validate the effectiveness of SCOUT, but also demonstrate the practical viability of Flow CoT as a scalable framework for enhancing reasoning in LLMs.

Large-scale wireless networks pose significant challenges in resource scheduling, where the solution space grows exponentially with network size. While network decomposition offers a promising solution by breaking networks into manageable subnetworks, existing approaches, including spectral clustering, fail to effectively capture the complex service relationships between base stations (BSs) and users, particularly in networks with massive user populations. This paper presents BSCCD (Bidirectional Spectral Co-Clustering Based Decomposition), a new decomposition scheme that addresses these challenges through two key innovations: (i) a two-round spectral co-clustering framework that captures bidirectional BS-user relationships, and (ii) a user node merging strategy that handles massive user populations. Extensive experiments on real-world datasets from multiple Chinese cities demonstrate that BSCCD reduces computation latency by up to 61.91 % compared to global optimization, while achieving more than 10 % improvement in solution quality over traditional clustering approaches. The advantage is particularly pronounced in medium-scale networks, where BSCCD outperforms traditional methods by$\mathbf{2 8. 7 6 \%}$. Our results demonstrate BSCCD's practical viability for resource scheduling in contemporary wireless networks, especially in scenarios with complex BS-user interactions and large user populations.

Deep learning-based Autonomous Driving (AD) perception models often exhibit poor generalization due to data heterogeneity in an ever domain-shifting environment. While Federated Learning (FL) could improve the generalization of an AD model (known as FedAD system), conventional models often struggle with under-fitting as the amount of accumulated training data progressively increases. To address this issue, instead of conventional small models, employing Large Vision Models (LVMs) in FedAD is a viable option for better learning of representations from a vast volume of data. However, implementing LVMs in FedAD introduces three challenges:(I)the extremely high communication overheads associated with transmitting LVMs between participating vehicles and a central server;(II)lack of computing resource to deploy LVMs on each vehicle;(III)the performance drop due to LVM focusing on shared features but overlooking local vehicle characteristics. To overcome these challenges, we propose pFedLVM, a LVM-Driven, Latent Feature-Based Personalized Federated Learning framework. In this approach, the LVM is deployed only on central server, which effectively alleviates the computational burden on individual vehicles. Furthermore, the exchange between central server and vehicles are the learned features rather than the LVM parameters, which significantly reduces communication overhead. In addition, we utilize both shared features from all participating vehicles and individual characteristics from each vehicle to establish a personalized learning mechanism. This enables each vehicle’s model to learn features from others while preserving its personalized characteristics, thereby outperforming globally shared models trained in general FL. As a demonstration of the proposed pFedLVM, this paper focuses on the semantic segmentation (SSeg) task. Extensive experiments demonstrate that pFedLVM outperforms the existing state-of-the-art approach by 18.47%, 25.60%, 51.03% and 14.19% in terms of mIoU, mF1, mPrecision and mRecall, respectively.

Street Scene Semantic Understanding (denoted as TriSU) is a complex task for autonomous driving (AD). However, inference model trained from data in a particular geographical region faces poor generalization when applied in other regions due to inter-city data domain-shift. Hierarchical Federated Learning (HFL) offers a potential solution for improving TriSU model generalization by collaborative privacy-preserving training over distributed datasets from different cities. Unfortunately, it suffers from slow convergence because the data from different cities are with disparate statistical properties. Going beyond existing HFL methods, we propose a Gaussian heterogeneous HFL algorithm (FedGau) to address inter-city data heterogeneity so that convergence can be accelerated. In the proposed FedGau algorithm, both single RGB image and RGB dataset are modelled as Gaussian distributions for aggregation weight design. This approach not only differentiates each RGB image by respective statistical distribution, but also exploits the statistics of dataset from each city in addition to the conventionally considered data volume. With the proposed approach, the convergence is accelerated by 35.5%-40.6% compared to existing state-of-the-art (SOTA) HFL methods. On the other hand, to reduce the involved communication resource, we further introduce a novel performance-aware adaptive resource scheduling (AdapRS) policy. Unlike the traditional static resource scheduling policy that exchanges a fixed number of models between two adjacent aggregations, AdapRS adjusts the number of model aggregation at different levels of HFL so that unnecessary communications are minimized. Extensive experiments demonstrate that AdapRS saves 29.65% communication overhead compared to conventional static resource scheduling policy while maintaining almost the same performance.

Decentralized federated learning (DFL) enables clients to train a neural network model in a device-to-device (D2D) manner without central coordination. In practical systems, DFL faces challenges due to dynamic topology changes, timevarying channel conditions, and limited computational capability of the clients. These factors can affect the learning performance and efficiency of DFL. To address the aforementioned challenges, in this paper, we propose a graph neural network (GNN)–based algorithm to minimize the total delay and energy consumption on training and improve the learning performance of DFL in D2D wireless networks. In our proposed GNN, a multihead graph attention mechanism is used to capture different features of clients and wireless channels. We design a neighbor selection module which enables each client to select a subset of its neighbors for the participation of model aggregation. We develop a decoder that enables each client to determine its transmit power and computational resource. Experimental results show that our proposed algorithm achieves a lower total delay and energy consumption on training when compared with five baseline schemes. Furthermore, by properly selecting a subset of neighbors for each client, our proposed algorithm achieves similar testing accuracy to the full participation scheme.

Artificial Intelligence-Generated Content (AIGC) has emerged as a transformative paradigm, enabling the autonomous creation of diverse content. By offloading model inference tasks to the network edge that is closer to mobile users (MUs), Mobile Edge Computing (MEC) has the potential to significantly enhance the performance of AIGC services. In practice, however, it is challenging to optimally manage MEC-empowered AIGC services, due to the lack of well-defined AIGC-specific metrics, as well as the dynamic workload and computation-intensive nature of AIGC services. In this paper, we first define a novel AIGC metric based on extensive real data experiments, and then study thejoint task offloading and resource allocationproblem in a generic MEC-empowered AIGC network, where MUs can offload model inference tasks to local or remote Base Stations (BSs), aiming at maximizing their Quality of Experience (QoE). The problem is challenging due to the fast and randomly changing of environments, as well as the necessity for real-time, asynchronous decision-making. To tackle these challenges, we propose two deep reinforcement learning algorithms based on the Proximal Policy Optimization (PPO) framework:Single-Layer PPO (SL-PPO)andMulti-Layer PPO (ML-PPO), designed for slow-changing and fast-changing environments, respectively. In the SL-PPO algorithm, both task offloading and resource allocation decisions are made simultaneously when tasks arrive. In the ML-PPO algorithm, the task offloading decision is made immediately when tasks arrive, while the resource allocation decision is deferred until tasks are scheduled for processing or transmission in the corresponding queues. Simulation results show that (i) both algorithms outperform existing methods in the literature, and can increase the average utility by up to 47% and 48.8%; (ii) both algorithms can effectively manage the trade-off between latency and energy consumption.

Mobile edge computing (MEC) is essential for next-generation mobile network applications that prioritize various performance metrics, including delays and energy efficiency. However, conventional single-objective scheduling solutions cannot be directly applied to practical systems in which the preferences (i.e., the weights of different objectives) are often unknown or challenging to specify in advance. In this study, we formulate a multi-objective offloading problem for MEC with multiple edges to minimize the sum of expected long-term energy consumption and delay while considering unknown preferences. To address the challenge of unknown preferences and the potentially diverse MEC systems, we propose a generalizable multi-objective (deep) reinforcement learning (GMORL)-based tasks offloading framework, which employs the Discrete Soft Actor-Critic (Discrete-SAC) method. Our method uses a single policy model to efficiently schedule tasks based on varying preferences and adapt to heterogeneous MEC systems with different CPU frequencies and server quantities. Under the proposed framework, we introduce a histogram-based state encoding method for constructing features for multiple edges in MEC systems, a sophisticated reward function for accurately computing the utilities of delay and energy consumption, and a novel neural network architecture for improving generalization. Simulation results demonstrate that our proposed GMORL scheme enhances the hypervolume of the Pareto front by up to 121.0% compared to benchmarks.

Mobile edge computing (MEC) is a promising paradigm for real-time applications with intensive computational needs (e.g., autonomous driving), as it can reduce the processing delay. In this work, we focus on the timeliness of computational-intensive updates, measured by Age-of-Information (AoI), and study how to jointly optimize the task updating and offloading policies for AoI with fractional form. Specifically, we consider edge load dynamics and formulate a task scheduling problem to minimize the expected time-average AoI. The uncertain edge load dynamics, the nature of the fractional objective, and hybrid continuous-discrete action space (due to the joint optimization) make this problem challenging and existing approaches not directly applicable. To this end, we propose a fractional reinforcement learning (RL) framework and prove its convergence. We further design a model-free fractional deep RL (DRL) algorithm, where each device makes scheduling decisions with the hybrid action space without knowing the system dynamics and decisions of other devices. Experimental results show that our proposed algorithms reduce the average AoI by up to 57.6% compared with several non-fractional benchmarks.

In federated learning (FL), distributed users collaboratively train a neural network model under the coordination of a central server. However, during the training process, clients often exhibit time-varying availability and have non-independent and non-identically distributed (non-IID) datasets. This results in system-induced bias, as models trained by the available clients do not accurately represent the entire population, which includes both available and unavailable clients. To address this bias, we propose a pricing-based incentive mechanism to encourage clients to adjust their availability. First, we model the strategic interaction among a large number of FL clients as a non-cooperative game under an arbitrary pricing scheme. We demonstrate that this game is a potential game, and its equilibrium can be found by solving an optimization problem. Second, based on equilibrium analysis, we derive an optimal pricing scheme for scenarios with a large client population. For general scenarios with any number of clients, we propose a bi-level optimization algorithm that utilizes Particle Swarm Optimization (PSO) to determine the optimal pricing scheme. This algorithm can effectively handles the intricate correlation between the equilibrium and pricing scheme. Our experimental results, based on real-world client availability datasets, highlight the effectiveness of our proposed incentive mechanism in mitigating system-induced bias, with improvements of up to 99.5% compared to the uniform pricing benchmark. Furthermore, this mechanism enhances the FL convergence rate by up to 3.43 times.

Street Scene Semantic Understanding (denoted as TriSU) is a crucial but complex task for world-wide distributed autonomous driving (AD) vehicles (e.g., Tesla). Its inference model faces poor generalization issue due to inter-city domain-shift. Hierarchical Federated Learning (HFL) offers a potential solution for improving TriSU model generalization, but suffers from slow convergence rate because of vehicles’ surrounding heterogeneity across cities. Going beyond existing HFL works that have deficient capabilities in complex tasks, we propose a rapid-converged heterogeneous HFL framework (FedRC) to address the inter-city data heterogeneity and accelerate HFL model convergence rate. In our proposed FedRC framework, both single RGB image and RGB dataset are modelled as Gaussian distributions in HFL aggregation weight design. This approach not only differentiates each RGB sample instead of typically equalizing them, but also considers both data volume and statistical properties rather than simply taking data quantity into consideration. Extensive experiments on the TriSU task using across-city datasets demonstrate that FedRC converges faster than the state-of-the-art benchmark by 38.7%, 37.5%, 35.5%, and 40.6% in terms of mIoU, mPrecision, mRecall, and mF1, respectively. Furthermore, qualitative evaluations in the CARLA simulation environment confirm that the proposed FedRC framework delivers top-tier performance.

Mobile-edge computing (MEC) is a promising computing scheme to support computation-intensive AI applications in vehicular networks, by enabling vehicles to offload computation tasks to edge computing servers deployed on road side units (RSUs) that approximate to them. In this work, we consider an MEC-enabled vehicular edge network (VEN), where each vehicle can offload tasks to edge/cloud computing servers via vehicle-to-infrastructure (V2I) links or to other end-vehicles via vehicle-to-vehicle (V2V) links. In such acloud-edge–endcollaborative offloading scenario, we focus on the joint task offloading, scheduling, and resource allocation problem for vehicles, which is challenging due to the online and asynchronous decision-making requirement for each task. To solve the problem, we propose aMultilayer deep reinforcement learning(DRL)-based approach, where each vehicle constructs and trains three modules to make different layers’ decisions: 1)Offloading Module(first layer), determining whether to offload each task, by using the dueling and double deepQ-network (D3QN) framework; 2)Scheduling Module(second layer), determining where and how to offload each task in the offloading queues, together with the transmission power, by using the parameterized deepQ-network (PDQN) framework; and 3)Computing Module(third layer), determining how much computing resource to be allocated for each task in the computation queues, by using classic optimization techniques. We provide the detailed algorithm design and perform extensive simulations to evaluate its performance. Simulation results show that our proposed algorithm outperforms the existing algorithms in the literature, and can reduce the average cost by 25.86%–72.51% and increase the average satisfaction rate by 3.48%–90.53%.

Federated learning (FL) is a distributed machine learning scheme that enables clients to train a shared global model without exchanging local data. In FL, the presence of label noise can severely reduce the accuracy of the global model. Although some recent works have focused on designing algorithms for label denoising, they ignored the important issue that clients may not apply costly label denoising strategies due to them being self-interested and having heterogeneous valuations on the model accuracy. To fill this gap, we model the clients’ strategic interactions as a novel label denoising game and determine the clients’ equilibrium strategies. We prove that the equilibrium outcome always leads to a lower global model accuracy than the socially optimal solution does. To motivate the clients’ efficient label denoising behaviors, we propose a penalty-based incentive mechanism and design the degree of penalty for punishing the clients’ undesired denoising behaviors, addressing the inaccurate noise rate detection in FL. We prove that our mechanism can achieve social efficiency, individual rationality, and weak budget balance. Numerical experiments on MNIST and CIFAR-10 show that as clients’ data become noisier, the gap between the equilibrium outcome and the socially optimal solution increases, verifying the necessity of an incentive mechanism. We empirically show that our proposed mechanism improves the model accuracy by up to 4.4% and incentivizes clients to achieve equilibrium strategies that are close to the socially optimal solution.

In cross-silo federated learning (FL), organizations cooperatively train a global model with their local datasets. However, some organizations may act as free riders such that they only contribute a small amount of resources but can obtain a high-accuracy global model. Meanwhile, some organizations can be business competitors, and they do not trust each other or any third-party entity. In this work, our goal is to design a framework that motivates efficient cooperation among organizations without the coordination of a central entity. To this end, we propose a blockchain-empowered incentive mechanism framework for cross-silo FL. Under this incentive mechanism framework, we develop a distributed algorithm that enables organizations to achieve social efficiency, individual rationality, and budget balance without private information of the organizations. Our proposed algorithm has a proven convergence guarantee and empirically achieves a higher convergence rate than a benchmark method. Moreover, we propose a transaction minimization algorithm to reduce the number of transactions made among organizations in the blockchain. This algorithm is proven to achieve a performance no worse than twice the minimum value. The experimental results in a testbed show that our proposed framework enables organizations to achieve social efficiency within a relatively short iterative process.

To enable AI-empowered Internet-of-things (AIoT) applications, it is crucial to achieve real-time data inference (e.g., prediction, control) at network edge. However, resource-constrained Internet-of-things devices (IoTDs) may be incapable of accomplishing those computation-intensive and latency-sensitive inference tasks. To address this issue, it is promising to incorporate mobile edge computing (MEC) systems and let IoTDs offload their inference tasks to edge servers that have already cached the associated neural network model required for inference. In this work, we take into account the limited storage and computing capacity of edge servers and formulate a neural network model caching problem for an MEC system with edge inference, in order to maximize the inference accuracy and reduce the task delay. To handle the exponential growth of signaling overhead and the learning difficulty under huge number of widely-deployed edge servers, we propose a cooperative mean field multi-agent reinforcement learning framework and a mean field actor-critic algorithm to solve the aforementioned problem. Simulation results show that our proposed algorithm outperforms several benchmarks, especially in large-scale edge networks.

Decentralized federated learning (DFL) enables clients to train a neural network model in a device-to-device (D2D) manner without central coordination. In practical systems, DFL faces challenges due to the dynamic topology changes, time-varying channel conditions, and limited computational capability of devices. These factors can affect the performance of DFL. To address the aforementioned challenges, in this paper, we propose a graph neural network (GNN)-based approach to minimize the total delay on training and improve the learning performance of DFL in D2D wireless networks. In our proposed approach, a multi-head graph attention mechanism is used to capture different features of clients and channels. We design a neighbor selection module which enables each client to select a subset of its neighbors for the participation of model aggregation. We develop a decoder which enables each client to determine its transmit power and CPU frequency. Experimental results show that our proposed algorithm can achieve a lower total delay on training when compared with three baseline schemes. Furthermore, the proposed algorithm achieves similar performance on the testing accuracy when compared with the full participation scheme.

Mobile edge computing (MEC) is essential for next-generation mobile network applications that prioritize various performance metrics, including delays and energy consumption. However, conventional single-objective scheduling solutions cannot be directly applied to practical systems in which the preferences of these applications (i.e., the weights of different objectives) are often unknown or challenging to specify in advance. In this study, we address this issue by formulating a multi-objective offloading problem for MEC with multiple edges to minimize expected long-term energy consumption and transmission delay while considering unknown preferences as parameters. To address the challenge of unknown preferences, we design a multi-objective (deep) reinforcement learning (MORL)-based resource scheduling scheme with proximal policy optimization (PPO). In addition, we introduce a well-designed state encoding method for constructing features for multiple edges in MEC systems, a sophisticated reward function for accurately computing the utilities of delay and energy consumption. Simulation results demonstrate that our proposed MORL scheme enhances the hypervolume of the Pareto front by up to 233.1% compared to benchmarks.

Information freshness, characterized by Age of Information (AoI), is important for sensor applications involving timely status updates. In many cases, the wireless signals from one sensor can be received by multiple access points (APs). This article investigates the average AoI for cooperative APs (Co-APs), in which they can share information through a wired backbone network. We first study a basic backbone-assisted Co-AP system where APs share only decoded packets. Experimental results on software-defined radios (SDRs) indicate that Co-AP significantly improves the average AoI performance over a single-AP system. Next, we investigate an improved Co-AP system, called Soft-Co-AP. In addition to sharing decoded packets, Soft-Co-AP shares and collects soft information of packets that the APs fail to decode for further joint decoding. A critical issue in Soft-Co-AP is determining the number of quantization bits that represent the soft information (each soft bit) shared over the backbone. While more quantization bits per soft bit improves the joint decoding performance, it leads to higher backbone delay. We experimentally study the average AoI of Soft-Co-AP by evaluating the tradeoff between the backbone delay and the number of quantization bits. SDR experiments show that when the number of sensors is large, Soft-Co-AP further reduces the average AoI by 12% compared with Co-AP. Interestingly, good average AoI performance is usually achieved when the number of quantization bits per soft bit is neither too large nor too small.

In mobile edge computing systems, an edge node may have a high load when a large number of mobile devices offload their tasks to it. Those offloaded tasks may experience large processing delay or even be dropped when their deadlines expire. Due to the uncertain load dynamics at the edge nodes, it is challenging for each device to determine its offloading decision (i.e., whether to offload or not, and which edge node it should offload its task to) in a decentralized manner. In this work, we consider non-divisible and delay-sensitive tasks as well as edge load dynamics, and formulate a task offloading problem to minimize the expected long-term cost. We propose a model-free deep reinforcement learning-based distributed algorithm, where each device can determine its offloading decision without knowing the task models and offloading decision of other devices. To improve the estimation of the long-term cost in the algorithm, we incorporate the long short-term memory (LSTM), dueling deep Q-network (DQN), and double-DQN techniques. Simulation results show that our proposed algorithm can better exploit the processing capacities of the edge nodes and significantly reduce the ratio of dropped tasks and average delay when compared with several existing algorithms.

360-degree video streaming provides users with immersive experience by letting users determine their field-of-views (FoVs) in real time. To efficiently utilize the limited bandwidth resources, recent works have proposed a viewport adaptive 360-degree video streaming model by exploiting the bitrate adaptation in spatial and temporal domains. In this paper, under this video streaming model, we propose an online bitrate selection algorithm to enhance the user’s quality of experience (QoE). This is achieved by characterizing the user’s personalized FoV and real-time downloading capacity in an online fashion. We address the unknown user-specific FoV by introducing the reference FoV and design an online bitrate selection algorithm to learn the difference between the user’s actual FoV and the reference FoV. We prove that as the number of video segments increases, the performance of the proposed online algorithm approaches the optimal performance asymptotically, with a bounded error. We perform trace-driven simulations with real-world datasets. Simulation results show that under the scenario where the available video bitrates are relatively high, our proposed algorithm can improve the user’s viewing quality level between$4.2\!-\!29.4$percent and reduce the average intra-segment quality switch by at least 12.4 percent when compared with several existing methods.

In cross-silo federated learning (FL), organizations cooperatively train a global model with their local data. The organizations, however, may be heterogeneous in terms of their valuation on the precision of the trained global model and their training cost. Meanwhile, the computational and communication resources of the organizations are non-excludable public goods. That is, even if an organization does not perform any local training, other organizations cannot prevent that organization from using the outcome of their resources (i.e., the trained global model). To address the organization heterogeneity and the public goods feature, in this paper, we formulate a social welfare maximization problem and propose an incentive mechanism for cross-silo FL. With the proposed mechanism, organizations can achieve not only social welfare maximization but also individual rationality and budget balance. Moreover, we propose a distributed algorithm that enables organizations to maximize the social welfare without knowing the valuation and cost of each other. Our simulations with MNIST dataset show that the proposed algorithm converges faster than a benchmark method. Furthermore, when organizations have higher valuation on precision, the proposed mechanism and algorithm are more beneficial in the sense that the organizations can achieve higher social welfare through participating in cross-silo FL.

Donation-based markets have been implemented by many online platforms, such as live streaming platforms. In these markets, producers provide services without mandatory charges, and customers enjoy the services and voluntarily donate money to the producers. The donation is split between the producers and platform with a pre-agreed fraction. To gain insights into the market operation, we use a two-stage game to capture the sequential decision process between the platform and producers. In Stage I, the platform decides a donation-split-fraction (DSF), i.e., the fraction of donation kept by the producers. In Stage II, producers decide whether to participate in the platform and (if yes) how to choose their service attributes considering the DSF as well as the producers' and customers' preferences. We prove that the Stage II game is a potential game with a counter-intuitive equilibrium result: although a larger DSF leads to more producer participation and a better match between the producers' choices and the customers' preferences, it does not necessarily lead to more total donation. The Stage I problem, nevertheless, is challenging to solve analytically due to its non-convexity. To gain insights regarding the optimal DSF that maximizes the platform's payoff, we characterize both its upper-bound and lower-bound. We show numerically that the platform's optimal payoff always decreases with the mismatch between the producers' and customers' preferences. Finally, we conduct a case study with the dataset from Twitch and demonstrate the approach of computing the platform's optimal DSF without the producers' inherent preferences.

Donation-based markets are becoming increasingly popular in our daily life. One example is the online streaming platform Twitch, which attracts millions of users on a daily basis. On such platforms, firms provide services to customers without mandatory charge, and customers voluntarily donate money to the firms. The donations are split between the firms and the platform with a fixed pre-agreed fraction. To gain insights into the operation and optimization of such platforms, we formulate a two-stage game to study the platform's and firms' behaviors. In Stage I, the platform decides a donation-split-fraction (DSF), which corresponds to the fraction of donations kept by the firms. In Stage II, firms decide whether to participate in the platform and how to choose their service attributes considering the DSF as well as the preferences of firms and customers. Analyzing such a two-stage game directly is challenging, as the Stage II problem corresponds to the multi-firm extension of the Hotelling model and is still an open problem. To resolve this issue, we approximate the large number of firms as non-atomic decision makers, where a single firm's strategy choice does not affect the payoffs of the firm population. Under such an approximation, we prove that the Stage II problem is a potential game. We further show that at the equilibrium, a larger DSF leads to more firm participations and a better match to the customers' preferences. The stage I problem, nevertheless, is a non-convex optimization problem that does not render a closed-form solution. To gain insights, we derive the upper-bound and lower-bound of the optimal DSF solution. The bounds suggest that the platform should increase its DSF if the customers' donation sensitivity to the number of firms increases or if the firms' opportunity cost for participation increases. Finally, we collect data from Twitch and demonstrate the results of the two-stage model with a case study. Our simulation results suggest that under our data and model settings, there exists a significant potential for Twitch to improve its payoff, by setting the DSF to 0.38, instead of 0.71 as in Twitch's current practice.

Adaptive bitrate streaming enables video users toadapttheir playing bitrates to the real-time network conditions, hence achieving the desirable quality-of-experience (QoE). In a multi-user wireless scenario, however, existing single-user based bitrate adaptation methods may fail to provide the desirable QoE, due to lack of consideration of multi-user interactions (such as the multi-user interferences and network congestion). In this work, we propose a novel user cooperation framework based onuser-provided networkingfor multi-user mobile video streaming over wireless cellular networks. The framework enables nearby mobile video users to crowdsource their cellular links and resources for cooperative video streaming. We first analyze the social welfare performance bound of the proposed cooperative streaming system by introducing a virtual time-slotted system. Then, we design a low complexity Lyapunov-based online algorithm, which can be implemented in an online and distributed manner without the complete future and global network information. Numerical results show that the proposed online algorithm achieves an average 97 percent of the theoretical maximum social welfare. We further conduct experiments with real data traces, to compare our proposed online algorithm with the existing online algorithms in the literature. Experiment results show that our algorithm outperforms the existing algorithms in terms of both the achievable bitrate (with an average gain of 20$\sim$30 percent) and social welfare (with an average gain of 10$\sim$50 percent).

Tactile Internet often requires: 1) the ultra-reliable and ultra-responsive network connection and 2) the proactive and intelligent actuation at edge devices. A promising approach to address these requirements is to enable mobile edge devices to share their communication, computation, and caching (3C) resources via device-to-device connections. In this paper, we propose a general 3C resource sharing framework, which includes many existing 1C/2C sharing models in the literature as special cases. Comparing with the 1C/2C models, the proposed 3C framework can further improve the resource utilization efficiency by offering more flexibilities in terms of the device cooperation and resource scheduling. As a typical example, we focus on the energy utilization under the proposed 3C framework. Specifically, we formulate an energy consumption minimization problem, which is an integer non-convex optimization problem. To solve the problem, we first transform it into an equivalent integer linear programming problem that is much easier to solve. Then, we propose a heuristic algorithm based on linear programming, which can further reduce the computation time and produce an empirically close-to-optimal solution. Moreover, we evaluate the energy reduction due to the 3C sharing both analytically and numerically. Numerical results show that, comparing with the existing 1C/2C approaches, the proposed 3C sharing framework can reduce the total energy consumption by 83.8% when the D2D energy is negligible. The energy reduction is still 27.5% when the D2D transmission energy per unit time is twice as large as the cellular transmission energy per unit time.

Today's mobile devices are capable of tackling various complicated tasks that may require a large amount of communication, computation, and caching (3C) resources. Due to users' heterogeneous resources and service requirements, it is challenging for each user to always accomplish his task satisfactorily. To alleviate this issue, mobile users can exploit the heterogeneity and crowdsource their resources to enhance the task execution performance. In this paper, we propose a general 3C framework that enables mobile users to share all three types of resources through device- to-device connections. Such a framework generalizes many existing 1C/2C resource sharing models (that only shares one or two types of resources among users). To quantify the benefit of the proposed framework, we focus on an energy minimization problem, and show that the 3C framework always achieves a smaller total energy consumption, comparing with other 1C/2C models. Furthermore, we show that the energy reduction is maximized, when user connection probability and content caching ratio are neither too large nor too small. Our numerical results show that, when ignoring device-to-device transmission energy, the general 3C framework can reduce the total energy consumption by 82.98%, comparing with the 1C/2C models.

Crowdsourced mobile video streaming enables nearby mobile video users to aggregate their network resources to improve the video streaming performance. However, users are often selfish and may not be willing to cooperate without proper incentives. Designing an incentive mechanism for such a scenario is challenging due to the users' asynchronous downloading behaviors as well as their private valuations for multi-bitrate encoded videos. In this work, we propose a multi-object multi-dimensional auction-based incentive framework, through which users can download multiple video segments with different bitrates for multiple nearby users (and themselves). Based on this incentive framework, we propose a Vickrey-score auction, which is the first multi-object multi-dimensional auction that achieves both truthfulness and efficiency. Simulations with real traces show that crowdsourced mobile streaming outperforms noncooperative streaming by 48.6% (on average) in terms of social welfare. We further implement our proposed auction mechanism in a demostration system, and show that the crowdsourced framework together with the auction mechanism can substantially increase mobile user's welfare and video service stability.

Accurate mobility prediction is becoming increasingly important in human behavior research, mainly due to many location-based applications such as mobile social networks and mobile advertisements. In this work, we propose a new crowd-sourced human mobility prediction model for public regions. We first analyze human trajectories collected through a cluster of densely deployed Wi-Fi access points (AP) in a shopping mall, and then characterize the close relationship between the human mobility patterns and the spatio-temporal contexts. Based on the distinct features of human trajectories in different types of public regions, we further propose a Markov-based crowdsourced mobility prediction method utilizing spatio-temporal contexts. We evaluate the performance of the proposed method using real traces, and show that our method is 28% more accurate in predicting human location transitions and incurs 14% smaller error in stay time prediction than the baseline methods.

Adaptive bitrate video streaming is a widely-used technology for mobile video streaming over HTTP. In this work, we study a crowdsourced video streaming framework, which enables nearby mobile users to crowdsource their radio resources for cooperatively adaptive bitrate video streaming. We propose a multi-dimensional auction based incentive mechanism to promote the user cooperation, supporting the asynchronous downloading and the bitrate adapting of video users. In this mechanism, each user initiates an auction whenever he is ready to download a new data segment in an asynchronous fashion, and all nearby users compete for the downloading opportunity by submitting a multidimensional bid consisting of the intended segment bitrate and the associated value. Design of such a multi-dimensional auction is very challenging, as we need to guarantee the user's truthful reporting on the information on multiple dependent dimensions. We first propose a truthful second-score (multi-dimensional) auction framework, within which we further derive the efficient mechanism that maximizes the social welfare (of each segment downloading) and the sub-optimal mechanism that approximately maximizes the auctioneer payoff. Experiment results show that our proposed crowdsourced streaming can achieve 60% ˜ 76% of the maximum social welfare even when 80 percentage of users lose their direct network connections.