Recent advancements in array-camera videography enable real-time capturing of ultra-high-definition (Ultra-HD) videos, providing rich visual information in a large field of view. However, promptly processing such data using state-of-the-art transformer-based vision foundation models faces significant computational overhead in on-device computing or transmission overhead in cloud computing. In this paper, we present Hyperion, the first cloud-device collaborative framework that enables low-latency inference on Ultra-HD vision data using off-the-shelf vision transformers over dynamic networks. Hyperion addresses the computational and transmission bottleneck of Ultra-HD vision Transformer inference by exploiting the intrinsic property in vision Transformer models. Specifically, Hyperion integrates a collaboration-aware importance scorer that identifies critical regions at the patch level, a dynamic scheduler that adaptively adjusts patch transmission quality to balance latency and accuracy under dynamic network conditions, and a weighted ensembler that fuses edge and cloud results to improve accuracy. Experimental results on real-world prototypes and datasets demonstrate that Hyperion enhances frame processing rate by up to 1.61 × and improves the accuracy by up to 20.2% when compared with state-of-the-art baselines under various network environments.

In today's cloud networks, host servers widely deploy Data Processing Units (DPUs) as network accelerators under the "Sep-Path" paradigm. However, as server capabilities scale with increasing CPU cores and network bandwidth, the software slow path (executed on a DPU's CPU) has become a critical bottleneck for workloads with high new-flow rates. Meanwhile, new-flow setup logic on host servers must continuously evolve to meet diverse and changing customer demands, making flexibility a key requirement alongside performance. To address this gap, we present XFir, the first hardware-accelerated new-flow setup system for cloud host servers that delivers high CPS throughput while preserving sufficient flexibility. XFir leverages a next-generation DPU equipped with a Cloud Network co-Processor (CNP) to execute the host server's new-flow setup logic. XFir redesigns the host-server flow-setup datapath and table layout, optimizes LPM lookups, and introduces CPU-CNP collaboration mechanisms to further improve performance and reliability. Our evaluation shows that XFir achieves over 776K new-flow CPS on a single host server with 11.7μs slow-path latency. Compared to prior work (Fornax), XFir achieves 4.8x CPS and reduces latency by 69.2%. Moreover, XFir is cost-effective to deploy, requiring only a single DPU per host. Overall, XFir improves new-flow throughput while maintaining development flexibility at low financial cost.

The rapid evolution of Large Language Models (LLMs) into multi-modal LLM agents has enabled autonomous systems capable of complex reasoning and action. While deploying these Multi-Agent Systems (MAS) at the edge promises reduced latency and enhanced privacy, it introduces significant challenges due to the heterogeneity of edge resources, the massive computational overhead of multimodal LLMs, and the volatility of multimodal data transmission. Existing service placement strategies often overlook the intricate dependencies within agent workflows and the substantial configuration latency required for model switching. In this paper, we propose a latency-aware algorithm for placing heterogeneous multimodal LLM agents at the edge. We model agent collaboration as a Directed Acyclic Graph (DAG), and then formulate the placement problem to maximize the number of satisfied user requests under strict resource constraints. We introduce a dynamic programming-based algorithm with theoretical performance bounds for single-request optimization, explicitly accounting for dynamic model configuration costs. For multi-request scenarios, we develop an online list-scheduling algorithm leveraging B-level heuristics to maximize request completion rates. Extensive experiments using five state-of-the-art agent systems (e.g., AWorld, OpenManus) across four benchmarks (e.g., GAIA, OSWorld) demonstrate that our approach reduces average latency by up to 42.62% compared to baselines, and therefore significantly improves throughput in high-load environments.

Applications based on Large Language Models (LLMs) contain a series of tasks to address real-world problems with boosted capability, which have dynamic demand volumes on diverse backends. Existing serving systems treat the resource demands of LLM applications as a blackbox, compromising end-to-end efficiency due to improper queuing order and backend warm up latency. We find that the resource demands of LLM applications can be modeled in a general and accurate manner with Probabilistic Demand Graph (PDGraph). We then propose Hermes, which leverages PDGraph for efficient serving of LLM applications. Confronting probabilistic demand description, Hermes applies the Gittins policy to determine the scheduling order that can minimize the average application completion time. It also uses the PDGraph model to help prewarm cold backends at proper moments. Experiments with diverse LLM applications confirm that Hermes can effectively improve the application serving efficiency, reducing the average completion time by over 70% and the P95 completion time by over 80%.

Neural Radiance Fields (NeRF) offers superior reconstruction accuracy over traditional methods like point clouds. This is particularly advantageous for applications in retail, VR, and navigation. However, when representing a complex scene with multiple semantically meaningful objects, especially those containing high-frequency details, the conventional single NeRF approach struggles to accurately capture these intricate details. Furthermore, the state-of-the-art multi-NeRF solutions, though effectively capturing details by representing each object as a separate NeRF, are too resource-intensive to be run on mobile devices. Existing NeRF rendering acceleration studies focus on single NeRF data representation and its framework improvements, without resorting to the visibility information for further optimization and considering the practical multi-NeRF rendering scenario. In this paper, we present Mosaic, a multi-NeRF cloud rendering system that facilitates high-fidelity, and low-latency cloud rendering for complex scenes. We first design a multi-resolution controller to dynamically determine the resolution for each pixel based on its contextual details and view direction. It then incorporates a visibility-aware region pruner to effectively eliminate unneeded pixels to avoid being rendered. The extensive experiments on various scenes show that Mosaic can achieve high-fidelity rendering with 43% lower transmission cost and up to 70% smaller computation consumption when compared with other state-of-the-art benchmarks.

Federated Learning (FL) allows edge clients to collaborate in model training with data privacy preserved, yet it is known to suffer low training efficiency and model accuracy. Given that efficiency and accuracy are usually conflicting objectives, existing practices increasingly employ an adaptive scheme that changes the FL configurations (e.g., quantization or sparsification level) based on runtime training status, for which accurate training diagnosis—used for guiding the optimization actions—is crucial. However, while training diagnosis is a common task shared by different optimization schemes, existing works propose their diagnosis methods in an ad-hoc manner, which yield multiple limitations. First, the diagnosis metric in an optimization scheme may sometimes be less accurate than others; second, existing schemes fail to fully exploit the diagnosis result by applying it for only one optimization action; third, existing methods usually do not perceive cross-client data heterogeneity, failing to simultaneously enhance FL accuracy. To tackle those limitations, we make a systematical study on the training diagnosis methods of multiple optimization schemes, and propose metric grafting—replacing a scheme's diagnosis metric with a better one to improve the training performance. Moreover, to fully exploit the potential of training diagnosis, we build a system platform that supports flexible combinations of training diagnosis and optimization actions (i.e., single-diagnosis-multiple actions and multiple-diagnosis-multiple-actions). Evaluation on testbeds show that, with metric grafting and advanced diagnosis action combinations, we can substantially improve the efficiency and accuracy performance of FL.

With the rapid development of edge devices and growing awareness of privacy protection, recent developers have transformed to fine-tune LLMs via federated learning instead of centralized training. Federated fine-tuning can leverage distributed data sources and computation power, but it also suffers from system and statistical heterogeneity. Client selection is an effective tool to solve the system and statistical heterogeneity in FL, but existing client selection schemes that involve online measurement will not be as effective in LLM fine-tuning as in conventional FL due to the huge LLM size and fewer fine-tuning rounds. In this paper, to the best of our knowledge, we are the first to consider both system and statistical heterogeneity in federated LLM fine-tuning, and we formulate a new latency minimization problem. We propose to measure client data overlap via knowledge graph offline to assist client selection in federated LLM fine-tuning. Our client selection scheme excels in both model accuracy and fine-tuning latency. We evaluate our scheme via two LLMs and two applications via four datasets. The experiment results illustrate that our scheme achieves the highest accuracy while 2.05x faster than the baselines.

Federated Learning (FL) is a distributed paradigm that supports collaborated model training while preserving data privacy, where clients periodically synchronize their local gradients once after multiple local iterations. Due to non-uniform data distribution and poor network condition, FL processes often suffer degraded training accuracy and efficiency. In this work, we analyze the microscopic parameter variation behaviors in FL, and find that an effective method to improve FL accuracy is to switch to more frequent synchronization at proper moments. In particular, such frequency-tuning moments—which can be detected from gradient characteristics—areheterogeneousacross different parameters. Motivated by such observations, we propose Parameter-Adaptive Synchronization (PAS), a FL scheme that adaptively tunes the synchronization period for each scalar parameter. The benefits of PAS are two-fold: By switching to more frequent synchronization when necessary, we can improve the FL training accuracy; by synchronizing different parameters independently, we can enable communication-computation overlapping and enhance the network utilization. We have theoretically demonstrated the convergence validity of PAS, and have further extended it with adaptive sparsification capability to jointly reduce the overall communication volume. We implemented PAS atop PyTorch, and extensive experiments show that it can substantially improve FL performance in both accuracy and communication efficiency.

Large Language Models (LLMs) have shown great promise in automating data analytics tasks by interpreting natural language queries and generating multi-operation execution plans. However, existing LLM-agent-based analytics frameworks operate under the assumption of centralized data access, offering little to no privacy protection. In contrast, federated analytics (FA) enables privacy-preserving computation across distributed data sources, but lacks support for natural language input and requires structured, machine-readable queries. In this work, we present LAFA, the first system that integrates LLM-agent-based data analytics with FA. LAFA introduces a hierarchical multi-agent architecture that accepts natural language queries and transforms them into optimized, executable FA workflows. A coarse-grained planner first decomposes complex queries into sub-queries, while a fine-grained planner maps each sub-query into a Directed Acyclic Graph of FA operations using prior structural knowledge. To improve execution efficiency, an optimizer agent rewrites and merges multiple DAGs, eliminating redundant operations and minimizing computational and communicational overhead. Our experiments demonstrate that LAFA consistently outperforms baseline prompting strategies by achieving higher execution plan success rates and reducing resource-intensive FA operations by a substantial margin. This work establishes a practical foundation for privacy-preserving, LLM-driven analytics that supports natural language input in the FA setting.

The proliferation of healthcare IoT devices and the resulting rich healthcare data sprout new possibilities for intelligent healthcare applications. Patients in intensive care units (ICUs) rely on various networked gadgets to continuously monitor their health and manage critical situations. Among the common therapeutic interventions in ICUs, invasive mechanical ventilation and injecting sedatives during ventilation play crucial roles in maintaining respiratory function and enhancing patient care. While existing therapeutic interventions largely depend on experience and intuition, we propose a federated inverse reinforcement learning framework, termed FERRY, which automatically and intelligently learns optimal therapeutic intervention policies across networked ICUs while keeping raw data local. Specifically, our federated approach overcomes limitations in medical data privacy and facilitates collaboration; our proposed inverse reinforcement learning framework learns the variational posterior distribution from historical trajectories to handle the unknown reward. Additionally, we enhance our framework with distributionally robust optimization to ensure worst-case performance and adaptively filter out noisy data through joint loss learning. Extensive experiments on the real-world dataset demonstrate that FERRY improves the overall ventilation and sedation decision-making accuracy by 36.75% compared to other state-of-the-art baselines.

Neural Radiance Fields (NeRF) is a cutting-edge neural network-based technique for novel view synthesis in 3D reconstruction. However, its significant computational demands pose challenges for deployment on mobile devices. While mesh-based NeRF solutions have shown potential in achieving real-time rendering on mobile platforms, they often fail to deliver high-quality reconstructions when rendering practical complex scenes. Additionally, the non-negligible memory overhead caused by pre-computed intermediate results complicates their practical application. To overcome these challenges, we present NeRFlex, a resource-aware, high-resolution, real-time rendering framework for complex scenes on mobile devices. NeRFlex integrates mobile NeRF rendering with multi-NeRF representations that decompose a scene into multiple sub-scenes, each represented by an individual NeRF network. Crucially, NeRFlex considers both memory and computation constraints as first-class citizens and redesigns the reconstruction process accordingly. NeRFlex first designs a detail-oriented segmentation module to identify sub-scenes with high-frequency details. For each NeRF network, a lightweight profiler, built on domain knowledge, is used to accurately map configurations to visual quality and memory usage. Based on these insights and the resource constraints on mobile devices, NeRFlex presents a dynamic programming algorithm to efficiently determine configurations for all NeRF representations, despite the NP-hardness of the original decision problem. Extensive experiments on real-world datasets and mobile devices demonstrate that NeRFlex achieves real-time, high-quality rendering on commercial mobile devices.

Developing compound Large Language Model (LLM) applications is becoming an increasingly prevalent approach to solving real-world problems. In these applications, an LLM collaborates with various external modules, including APIs and even other LLMs, to realize complex intelligent services. However, we reveal that the intrinsic duration and structural uncertainty in compound LLM applications pose great challenges for LLM service providers in serving and scheduling them efficiently. In this paper, we propose LLMSched, an uncertainty-aware scheduling framework for emerging compound LLM applications. In LLMSched, we first design a novel DAG-based model to describe the uncertain compound LLM applications. Then, we adopt the Bayesian network to comprehensively profile compound LLM applications and identify uncertainty-reducing stages, along with an entropy-based mechanism to quantify their uncertainty reduction. Combining an uncertainty reduction strategy and a job completion time (JCT)-efficient scheme, we further propose an efficient scheduler to reduce the average JCT. Evaluation of both simulation and testbed experiments on various representative compound LLM applications shows that compared to existing state-of-the-art scheduling schemes, LLMSched can reduce the average JCT by 14 ~ 79%.

Vision Transformers (ViTs) have outperformed traditional Convolutional Neural Network architectures and achieved state-of-the-art results in various computer vision tasks. Since ViTs are computationally expensive, the models either have to be pruned to run on resource-limited edge devices only or have to be executed on remote cloud servers after receiving the raw data transmitted over fluctuating networks. The resulting degraded performance or high latency all hinder their widespread applications. In this paper, we present Janus, the first framework for low-latency cloud-device collaborative Vision Transformer inference over dynamic networks. Janus overcomes the intrinsic model limitations of ViTs and realizes collaboratively executing ViT models on both cloud and edge devices, achieving low latency, high accuracy, and low communication overhead. Specifically, Janus judiciously combines token pruning techniques with a carefully designed fine-to-coarse model splitting policy and non-static mixed pruning policy. It attains a balance between accuracy and latency by dynamically selecting the optimal pruning level and split point. Experimental results across various tasks demonstrate that Janus enhances throughput by up to 5.15x and reduces latency violation ratios by up to 98.7% when compared with baseline approaches under various network environments.

Low Earth Orbit (LEO) satellite networks, such as Starlink, have become essential complements to terrestrial networks, significantly expanding Internet access in rural areas. However, their robustness in the face of various threats remains largely unexplored. In this paper, we present the first comprehensive study on the impact of four critical factors—direct sunlight, solar superstorms, high workload, and wear and tear—on Inter-Satellite Link (ISL) failures and their effects on network routing performance. We systematically model these failure mechanisms and analyze their influence under three classical routing algorithms. Our evaluation includes both network-wide connectivity and end-user latency analysis. Our findings reveal that LEO satellite networks relying on static routing algorithms are highly vulnerable, whereas dynamic and distributed routing approaches exhibit greater resilience. While multipath routing enhances robustness through redundancy, it encounters two key challenges: throughput bottlenecks and performance degradation under high-latitude traffic loads. In contrast, distributed routing maintains resilience with minimal latency trade-offs. These insights underscore the need for adaptive routing frameworks that integrate spatial redundancy with traffic-aware load balancing to ensure reliable network performance. Code for this paper is available at https://github.com/zpatronus/Robustness-Analysis-of-LEO-Satellite-Routing.

Existing quality of experience (QoE)-driven adaptive bitrate (ABR) algorithms either fail to consider personalized QoE or rely on over-simplified QoE models, all resulting in unsatisfactory streaming experiences. Recognizing the wide existence of user feedback schemes in existing streaming applications, we introduce Q+, a framework leveraging progressively gathered personal user opinion scores from multiple interaction sessions for enhanced user-system alignment. Q+ first innovates QoE modeling by incorporating both pairwise ordinal and cardinal preferences constructed from scores. The capturing of both preferences ensures reliable and robust preference representation. Moreover, we design a monotonic neural network as the QoE model to capture the inherent monotonicity property in ABR services, improving model expressivity and generalization ability even with limited human feedback. To align the policy with the progressively updated QoE, we then develop a value-based reinforcement learning (RL) algorithm for bitrate control that integrates reward relabeling and calibrated prioritized experience replay. Extensive experiments reveal that Q+ consistently surpasses state-of-the-art rule-based, control-based, and RL-based baselines within only three sessions, improving QoE by 5.69% to 29.39% across diverse network conditions.

With the rapid growth of Low Earth Orbit (LEO) satellite networks, satellite-IoT systems using the LoRa technique have been increasingly deployed to provide widespread Internet services to low-power and low-cost ground devices. However, the long transmission distance and adverse environments from IoT satellites to ground devices pose a huge challenge to link reliability, as evidenced by the measurement results based on our real-world setup. In this paper, we propose a blind coherent combining design named B2LoRa to boost LoRa transmission performance. The intuition behind B2LoRa is to leverage the repeated broadcasting mechanism inherent in satellite-IoT systems to achieve coherent combining under the low-power and low-cost constraints, where each re-transmission at different times is regarded as the same packet transmitted from different antenna elements within an antenna array. Then, the problem is translated into aligning these packets at a fine granularity despite the time, frequency, and phase offsets between packets in the case of frequent packet loss. To overcome this challenge, we present three designs — joint packet sniffing, frequency shift alignment, and phase drift mitigation to deal with ultra-low SNRs and Doppler shifts featured in satellite-IoT systems, respectively. Finally, experiment results based on our real-world deployments demonstrate the high efficiency of B2LoRa.

Real-time industrial video analytics is widely applied across diverse domains within cyber-physical systems (CPS). CPS devices equipped with networked cameras are wirelessly connected to servers for complex vision-based analytics and intelligent operations. Adaptive video streaming is a pivotal technique in these applications to effectively deliver video content to servers under varying network conditions, enabling complex analytics afterward. Our thorough data analysis reveals that conventional offline video streaming control policies cannot effectively adapt to the high dynamics in networks and industrial video scenes. This results in suboptimal analytic performance and necessitates online adaptation for streaming control policy models. Yet, updating control policy models requires ground-truth analytics results which are unavailable directly on end devices due to their limited capacity. Furthermore, naively streaming original videos to the server for online adaptation is greatly challenged by scarce and dynamic networks, leading to decreased accuracy and increased transmission costs. In this paper, we present RIVA, a novel Online Learning-enabled adaptive streaming framework for Real-time Industrial Video Analytics. To facilitate communication-efficient online retraining, we design a hierarchical reinforcement learning approach in which the upper-level module intelligently determines the timing for online retraining, balancing Quality of Service (QoS) improvement and communication cost. Meanwhile, the lower-level module dynamically allocates bitrate to maximize QoS. Extensive experiments based on real-world industrial video and network datasets demonstrate that our proposed framework achieves a 22.6% mean accuracy increase, a 64.9% decrease in the mean failure rate of video uploading, and a 60.2% mean latency decrease compared to the state-of-the-art solutions.

Federated multi-armed bandits extend the multi-armed bandits framework to the federated learning setting where multiple clients in the same exploration space collaboratively identify the optimal arm. While previous studies demonstrated its efficiency like classical federated learning, in this paper, we present the first work that reveals the serious fairness problem in federated multi-armed bandits when clients have heterogeneous and overlapping armsets. The fairness problem happens because clients with different trial requirements should conduct different numbers of trials summing up to a global requirement, but they wish to minimize their exploration effort. To address the novel fairness concern, we formally formulate the fairness-aware trial assignment as a coalitional game. Based on the theoretically derived trial requirements, we devise a Shapley value-based trial assignment mechanism to guarantee fairness. Regardless of the #P-hard complexity when deriving the general Shapley value, we achieve an accurate computation of trial assignment with polynomial complexity by exploiting its unique characteristic. We further carefully control the numbers of trials in each iteration to resolve the communication bottleneck and minimize the wasted trials. Experiment results show that, compared to the naïve federated scheme, our design outperforms with both high fairness metrics and high efficiency in total trials and communication.

Tile-based streaming and super resolution (SR) are two representative technologies adopted to improve bandwidth efficiency of 360° video streaming. The former allows selective downloading of contents in the user viewport by splitting the video into multiple independently decodable tiles. The latter leverages client-side computation to enhance the received video to higher quality using advanced neural network models. In this work, we propose a Collaborated Streaming and Enhancement (CSE) adaptation framework for mobile 360° videos, which integrates super resolution with tile-based streaming to optimize the user experience with dynamic bandwidth and limited computing capability. To effectively enhance the tile-based video streaming through SR, we propose to adaptively group the tiles for quality enhancement adapting to the content similarity. We also identify and address several key design issues to integrate SR into tile-based video streaming including unified video quality assessment, computational complexity model for super resolution, and buffer analysis considering the interplay between transmission and enhancement. We further formulate the quality-of-experience (QoE) maximization problem for mobile 360° video streaming and propose a rate adaptation algorithm to make the best decisions for download and for enhancement based on the Lyapunov optimization theory. Extensive evaluation results validate the superiority of our proposed approach, which demonstrates stable performance with considerable QoE improvement, while enabling a trade-off between playback smoothness and video quality.

Due to its advantage of protecting data privacy and reducing communication overhead, Federated Learning (FL) is becoming a promising machine learning paradigm. However, resource limitations and unstable communication connections on the participating client end can lead to unintentional failures that degrade FL performance. Moreover, as FL systems scale and interconnect increasingly, they face growing exposure to intentional network risks. Furthermore, the assumption of continued trust in historically benign clients introduces vulnerabilities to potential internal betrayal within FL systems. In this paper, we enhance the robustness of FL by incorporating the zero-trust principle, which eliminates implicit trust in clients and mitigates unintentional failures, intentional attacks, and strategic betrayal risks. The framework incorporates dynamic client selection and aggregation weight allocation through trustworthiness evaluation and sustained skepticism toward each potential betrayal behavior. Specifically, a Dirichlet-based trust evaluation technique is presented to update clients' trustworthiness with evolving observations. Then, to reduce potential betrayal loss, we formulate a min-max optimization problem that minimizes the worst-case betrayal loss. Next, we transform the formulation into a convex programming problem for solution. Extensive simulations are conducted to demonstrate the efficacy of the zero-trust based FL in the accurate trust assessment and the system's betrayal-aware robustness enhancement.

The advent of deep learning has precipitated a surge in public machine learning as a service (MLaaS) for multimedia analysis. However, reliance on a single MLaaS can result in product dependency and a loss of better performance offered by multiple MLaaSes. Consequently, many enterprises opt for an intercloud broker capable of managing jobs across various clouds. Though existing works explore the efficient utilization of inter-cloud computational resources and the enhancement of inter-cloud data transfer throughput, they disregard improving the overall accuracy of multiple MLaaSes. In response, we conduct a measurement study on object detection services, which are designed to identify and locate various objects within an image. We discover that combining predictions from multiple MLaaSes can improve analytical performance. However, more MLaaSes do not necessarily equate to better performance. Therefore, we propose SkyML, a user-side MLaaS federation broker that selects a subset of MLaaSes based on the characteristics of the request to achieve optimal multimedia analytical performance. Initially, we design a combinatorial reinforcement learning approach to select the sound MLaaS combination, thereby maximizing user experience. We also present an ingenious, automated taxonomy unification algorithm to minimize human efforts in merging MLaaS-specific labels into a user-preferred label space. Moreover, we devise an optimized ensemble strategy to aggregate predictions from the selected MLaaSes. Evaluations indicate that our similarity-based taxonomy unification approach can reduce annotation costs by 90%. Moreover, real-world trace-driven evaluations further prove that our MLaaS selection method can achieve similar levels of accuracy with a 67% reduction in inference fees.

Recent years have witnessed a rapid growth of both 5G coverage and 5G users, attracting several measurement studies on its coverage, reliability and quality of service. However, the capabilities and potential impacts of 5G, especially Standalone (SA) 5G, still remain to be fully understood from a content provider (CP)’s perspective. This paper fills this gap by studying 5G networks used by over 23 million users in one year in Kuaishou, a popular crowdsourced live streaming platform. With passive and active measurements, we have the following key findings: i) SA 5G generally provides end-to-end performance improvements compared to 4G or Non-Standalone (NSA) 5G, but its advantage depends on both the number of cellular users and CP-level configurations. ii) In the radio access networks, SA 5G is more sensitive to access density, but has better handover tolerance. iii) Controlled experiments with 29 mobile device models on energy consumption refute some “conventional wisdom,” including the notion that 5G always consumes more power. iv) Traceroute-based active experiments in over 300 cities show that although users are “closer” to the internet in SA 5G due to the control and user plane separation, whether end-to-end latency benefits from that partly depends on the routing policy at the gateways. Furthermore, we propose a 5G-aware rebuffer strategy tested by 9 million viewers in Kuaishou, showing a 7% reduction in rebuffer proportion. Finally, we also provide new design space for other 5G participants.

Serverless computing has become increasingly popular due to its flexible and hassle-free service, relieving users from traditional resource management burdens. However, the shift in responsibility has led to unprecedented challenges for serverless providers in managing virtual machines (VMs) and serving heterogeneous function instances. Serverless providers need to purchase, provision and manage VM instances from IaaS providers, aiming to minimize VM provisioning costs while ensuring compliance with Service Level Objectives (SLOs). In this paper, we propose Trident, a provider-oriented resource management framework for serverless computing platforms. Trident optimizes three major serverless computing provisioning problems for serverless providers: workload prediction, VM provisioning, and function placement. Specifically, Trident introduces a novel dynamic model selection algorithm for more accurate workload prediction. With the prediction results, Trident then carefully designs a hierarchical reinforcement learning (HRL)-based approach for VM provisioning with a mix of types and configurations. To further improve resource utilization, Trident employs an effective collocation placement strategy for efficient function container scheduling. Evaluations on the Azure Function dataset demonstrate that Trident maintains the lowest probability of violating SLOs while simultaneously achieving substantial cost savings of up to 71.8% in provisioning expense compared to state-of-the-art methods from industry and academia.

Nowadays, Federated Learning (FL) has emerged as a promising and critical machine learning scheme to protect data privacy and reduce communication overhead. As the scale and connectivity expand in the FL system, enhancing the model’s robustness against security threats from malicious clients grows ever more critical. An effective defensive solution involves selecting benign clients appropriately, thereby mitigating the vulnerability of the FL system to malicious attacks. However, clients exhibit varying behaviors over time, which complicates the task of accurately modeling their future trustworthiness. Moreover, blindly trusting clients with high trust values poses risks, given the potential for severe losses from betrayal. To tackle these problems, we propose a zero-trust policy in FL aimed at establishing continuous trust in each client while maintaining skepticism towards potential betrayal attacks. Specifically, we develop a Dirichlet-based trust evaluation technique to enable a comprehensive selection of trustworthy participants. This technique leverages the posterior distribution to estimate clients’ trust values from their evolving behavior records over time. Then, we anticipate potential betrayal from a selected client and formulate a min-max optimization problem to minimize the worst-case betrayal loss, thereby boosting the system’s betrayalaware robustness. Next, we convert this problem into a convex optimization problem and utilize the interior point method for resolution. We conduct extensive simulations to validate the efficacy of our proposed zero-trust policy in accurately assessing trust and enhancing the model’s robustness to betrayal.

Low-earth-orbit (LEO) satellite communication networks have evolved into mega-constellations with hundreds to thousands of satellites inter-connecting with inter-satellite links (ISLs). Network planning, which plans for network resources and architecture to improve the network performance and save operational costs, is crucial for satellite network management. However, due to the large scale of mega-constellations, high dynamics of satellites, and complex distribution of real-world traffic, it is extremely challenging to conduct scalable network planning on mega-constellations with high performance. In this paper, we propose SATFLOW, a distributed and hierarchical network planning framework to plan for the network topology, traffic allocation, and fine-grained ISL terminal power allocation for mega-constellations. To tackle the hardness of the original problem, we decompose the grand problem into two hierarchical sub-problems, tackled by two-tier modules. A multi-agent reinforcement learning approach is proposed for the upper-level module so that the overall laser energy consumption and ISL operational costs can be minimized; A distributed alternating step algorithm is proposed for the lower-level module so that the laser energy consumption could be minimized with low time complexity for a given topology. Extensive simulations on various mega-constellations validate SATFLOW's scalability on the constellation size, reducing the flow violation ratio by up to$21.0 \%$and reducing the total costs by up to$89.4 \%$, compared with various state-of-the-art benchmarks.

Large constellations of Earth Observation Low Earth Orbit satellites collect enormous amounts of image data every day. This amount of data needs to be transferred to data centers for processing via ground stations. Ground Station as a Service (GSaaS) emerges as a new cloud service to offer satellite operators easy access to a network of ground stations on a pay-per-use basis. However, renting ground station and data center resources still incurs considerable costs, especially for large satellite constellations. The current practice of sticking to a single GSaaS provider also suffers high data latency and low robustness to weather variability due to limited ground station availability. To address these limitations, we propose SkyGS, a system that schedules both communication and computation by federating GSaaS and cloud computing services across multiple cloud providers. We formulate the resulting problem as a system cost minimization problem with a long-term data latency threshold constraint. In SkyGS, we apply Lyapunov optimization to decompose the long-term optimization problem into a series of real-time optimization problems that do not require prior knowledge. As the decomposed problem is still of exponential complexity, we transform it into a bipartite graph-matching problem and employ the Hungarian algorithm to solve it. We analyze the performance theoretically and evaluate SkyGS using realistic simulations based on real-world satellites, ground stations, and data centers data. The comprehensive experiments demonstrate that SkyGS can achieve cost savings by up to 63% & reduce average data latency by up to 95%.

Federated learning (FL) has emerged as a prevalent distributed machine learning scheme that enables collaborative model training without aggregating raw data. Cloud service providers further embrace Federated Learning as a Service (FLaaS), allowing data analysts to execute their FL training pipelines over differentially-protected data. Due to the intrinsic properties of differential privacy, the enforced privacy level on data blocks can be viewed as a privacy budget that requires careful scheduling to cater to diverse training pipelines. Existing privacy budget scheduling studies prioritize either efficiency or fairness individually. In this paper, we propose DPBalance, a novel privacy budget scheduling mechanism that jointly optimizes both efficiency and fairness. We first develop a comprehensive utility function incorporating data analyst-level dominant shares and FL-specific performance metrics. A sequential allocation mechanism is then designed using the Lagrange multiplier method and effective greedy heuristics. We theoretically prove that DPBalance satisfies Pareto Efficiency, Sharing Incentive, Envy-Freeness, and Weak Strategy Proofness. We also theoretically prove the existence of a fairness-efficiency tradeoff in privacy budgeting. Extensive experiments demonstrate that DPBalance outperforms state-of-the-art solutions, achieving an average efficiency improvement of 1.44× ~ 3.49×, and an average fairness improvement of 1.37×~24.32×.

The emerging Web 3.0 paradigm aims to decentralize existing web services, enabling desirable properties such as transparency, incentives, and privacy preservation. However, current Web 3.0 applications supported by blockchain infrastructure still cannot support complex data analytics tasks in a scalable and privacy-preserving way. This paper introduces the emerging federated analytics (FA) paradigm into the realm of Web 3.0 services, enabling data to stay local while still contributing to complex web analytics tasks in a privacy-preserving way. We propose FedWeb, a tailored FA design for important frequent pattern mining tasks in Web 3.0. FedWeb remarkably reduces the number of required participating data owners to support privacy-preserving Web 3.0 data analytics based on a novel distributed differential privacy technique. The correctness of mining results is guaranteed by a theoretically rigid candidate filtering scheme based on Hoeffding’s inequality and Chebychev’s inequality. Two response budget saving solutions are proposed to further reduce participating data owners. Experiments on three representative Web 3.0 scenarios show that FedWeb can improve data utility by ∼25.3% and reduce the participating data owners by ∼98.4%.

Federated Learning (FL) is a distributed paradigm that supports collaborated model training while preserving data privacy, where clients periodically synchronize their local gradients once after multiple local iterations. Due to non-uniform data distribution and poor network condition, FL processes often suffer degraded training accuracy and efficiency. In this work, we analyze the microscopic parameter variation behaviors in FL, and find that an effective method to improve FL accuracy is to switch to more frequent synchronization at proper moments. Moreover, such moments can be detected from gradient characteristics, and are heterogeneous across different parameters. Motivated by such observations, we propose Parameter-Adaptive Synchronization (PAS), a FL scheme that adaptively tunes the synchronization period for each scalar parameter. The benefits of PAS are two-fold: By switching to more frequent synchronization when necessary, we can improve the FL training accuracy; by synchronizing different parameters independently, we can enable communication-computation overlapping and enhance the network utilization. We implemented PAS atop PyTorch, and extensive experiments show that it can substantially improve FL performance in both accuracy and communication efficiency.

Drone-sourced live video analytics has extensive applications across diverse domains. Adaptive video streaming is a pivotal technique in these applications that targets at effectively delivering video content to servers under varying network conditions, enabling complex analytics afterward. However, our thorough data analysis reveals that conventional offline video streaming policies cannot effectively adapt to highly fluctuating drone network environments and dynamic changes in aerial view scenes. This results in suboptimal analytic performance and necessitates online adaptation for policy models. Yet, obtaining ground-truth analytics results directly from drones is infeasible due to their limited capacity. Furthermore, naively streaming original videos to the server for online adaption is greatly challenged by the scarce and dynamic networks, leading to decreased accuracy performance and escalated transmission cost if not properly designed. In this paper, we present OAVS, a novel online learning-enabled adaptive streaming framework for drone-sourced video analytics. To facilitate cost-effective online retraining, we design a hierarchical reinforcement learning approach in which the upper-level module intelligently determines the timing for online retraining, balancing machine-perceived quality of experience (QoE) improvement and transmission cost. Meanwhile, the lower-level module dynamically allocates bitrate to maximize machine-perceived QoE. Extensive experiments based on real-world drone video and aerial network datasets demonstrate that our proposed framework achieves a 17.7% mean accuracy increase, a 37.5% decrease in the mean failure rate of video uploading, and a 5.2% mean latency decrease compared to state-of-the-art solutions.

Utilizing compound Large Language Model (LLM) systems, instead of a monolithic LLM model, is gradually becoming a practical solution to realize a diverse range of industry applications. In compound LLM systems, an LLM collaborates with other external tools, APIs, or LLMs to offer intelligent services. In this poster, we identify the unique challenges, namely temporal and topological uncertainty, brought about by compound LLM systems in system serving. We then propose a priority-based scheduling policy to schedule different stages in DAG-represented compound LLM systems. The preliminary results show promising performance of uncertainty-aware scheduling policies.

Neural-based super-resolution (SR) has achieved great success in enhancing image or video quality, creating new opportunities for building bandwidth-efficient and high-accuracy video analytics (VAs) systems. Intuitively, with the help of SR techniques, cameras only need to send downsampled low-quality frames to the server in a canonical edge-assisted VAs framework. The server-side SR model then upscales the quality of received frames for the subsequent VAs tasks, incurring thus substantially reduced bandwidth consumption. Nonetheless, as revealed by our measurement results on real-world video clips, higher delivery quality does not necessarily lead to higher analysis accuracy. This motivates us to study the content-adaptive downsampling and upscaling ratio selection problem for VAs streaming. We propose an SR-based VAs framework, named AdaDSR that can dynamically select the optimal downsampling and upscaling ratios so that the system utility can be maximized. AdaSDR is configured to balance the tradeoffs among accuracy, network cost, and computational cost. It further leverages the temporal consistency of videos to skip trivial decisions so that the camera’s processing overhead can be reduced. Experiments on real-world video data sets demonstrate that AdaDSR can improve the average utility by 7.2%–18.4% when compared with state-of-the-art approaches under diverse video scenes.

Federated learning (FL) has emerged as a prominent distributed learning paradigm, enabling collaborative training of neural network models across local devices with raw data stay local. However, FL systems often encounter significant challenges due to data heterogeneity. Specifically, the non-IID dataset in FL systems substantially slows down the convergence speed during training and adversely impacts the accuracy of the final model. In our paper, we introduce a novel client selection framework that judiciously leverages correlations across local datasets to accelerate training. Our framework first employs a lightweight locality-sensitive hashing algorithm to extract client features while respecting data privacy and incurring minimal overhead. We then design a novel Neural Contextual Combinatorial Bandit (NCCB) algorithm to establish relationships between client features and rewards, enabling intelligent selection of client combinations. We theoretically prove that our proposed NCCB has a bounded regret. Extensive experiments on real-world datasets further demonstrate that our framework surpasses state-of-the-art solutions, resulting in a 50% reduction in training time and a 17% increase in final model accuracy, closing to the performance in the ideal IID case.

Network traffic classifiers of mobile devices are widely learned with federated learning(FL) for privacy preservation. Noisy labels commonly occur in each device and deteriorate the accuracy of the learned network traffic classifier. Existing noise elimination approaches attempt to solve this by detecting and removing noisy labeled data before training. However, they may lead to poor performance of the learned classifier, as the remaining traffic data in each device is few after noise removal. Motivated by the observation that the data feature of the noisy labeled traffic data is clean and the underlying true distribution of the noisy labeled data is statistically close to the clean traffic data, we propose to utilize the noisy labeled data by normalizing it to be close to the clean traffic data distribution. Specifically, we first formulate a distributionally robust federated network traffic classifier learning problem (DR-NTC) to jointly take the normalized traffic data and clean data into training. Then we specify the normalization function under Wasserstein distance to transform the noisy labeled traffic data into a certified robust region around the clean data distribution, and we reformulate the DR-NTC problem into an equivalent DR-NTC-W problem. Finally, we design a robust federated network traffic classifier learning algorithm, RFNTC, to solve the DR-NTC-W problem. Theoretical analysis shows the robustness guarantee of RFNTC. We evaluate the algorithm by training classifiers on a real-world dataset. Our experimental results show that RFNTC significantly improves the accuracy of the learned classifier by up to 1.05 times.

High Definition (HD) maps have become core supporting components for autonomous driving. To date, their updates heavily depend on the vehicle fleets of the map vendors, which cannot scale and timely reflect the highly dynamic environment. To ensure the HD map quality, it is advocated social vehicles should be used. Nevertheless, there are privacy concerns and a lack of incentives for social vehicles to contribute data. In this paper, we leverage federated analytics (FA), a newly developed collaborative data analytics paradigm, where raw data are kept local and only the insights generated from local analytics are sent to a server for aggregation. We present a new Federated Analytics based HD map Updating model (FAUMap) to protect the privacy of social vehicles. To motivate social vehicles to contribute data and improve the HD map quality, we formulate an overlapping coalition formation game, OCFUMap, and develop an algorithm to find feasible coalitions. Simulations show that our approach can improve the quality of the updated HD map by 1.56 times. To study an end-to-end operation of the FAUMap model and OCFUMap game, we present a case of HD map updates of the Powell street in San Francisco using the autonomous driving simulator CarLA.

The advancement of nanosatellite techniques has boosted the growth of satellite-originated data and applications. Satellite edge computing (SEC) is envisioned to provide in-orbit processing of the sensed data to save the scarce terrestrial-satellite communication resources and support mission-critical services. While most of the existing SEC studies mainly focus on general computing tasks, we present a two-tier collaborative processing framework for the important and unique hyperspectral image (HSI) processing task. Our framework carefully selects bands out of the collected HSIs and sends them back for further analysis. We first conduct a comprehensive data analysis to reveal the non-trivial relationship between the band selection and the eventual analytic performance. We then formulate the band selection problem in this collaborative setting as a utility maximization problem that jointly considers the analytic, energy, and communication factors. A novel multi-agent reinforcement learning approach, named MaHSI, is proposed to solve it in the dynamic SEC environment. Our multi-agent design judiciously embeds the complex correlations among bands as collaborations among agents and significantly reduces the exploration space. Extensive experiments on real-world HSI datasets prove that our approach not only outperforms the existing classical band selection algorithms in accuracy and inference speed but also brings the highest utility to the satellites.

Graph neural networks (GNN) have been widely deployed in real-world networked applications and systems due to their capability to handle graph-structured data. However, the growing awareness of data privacy severely challenges the traditional centralized model training paradigm, where a server holds all the graph information. Federated learning is an emerging collaborative computing paradigm that allows model training without data centralization. Existing federated GNN studies mainly focus on systems where clients hold distinctive graphs or sub-graphs. The practical node-level federated situation, where each client is only aware of its direct neighbors, has yet to be studied. In this paper, we propose the first federated GNN framework called Lumos that supports supervised and unsupervised learning with feature and degree protection on node-level federated graphs. We first design a tree constructor to improve the representation capability given the limited structural information. We further present a Monte Carlo Markov Chain-based algorithm to mitigate the workload imbalance caused by degree heterogeneity with theoretically-guaranteed performance. Based on the constructed tree for each client, a decentralized tree-based GNN trainer is proposed to support versatile training. Extensive experiments demonstrate that Lumos outperforms the baseline with significantly higher accuracy and greatly reduced communication cost and training time.

With the reduced hardware costs of omnidirectional cameras and the proliferation of various extended reality applications, more and more 360° videos are being captured. To fully unleash their potential, advanced video analytics is expected to extract actionable insights and situational knowledge without blind spots from the videos. In this paper, we present OmniSense, a novel edge-assisted framework for online immersive video analytics. OmniSense achieves both low latency and high accuracy, combating the significant computation and network resource challenges of analyzing 360° videos. Motivated by our measurement insights into 360° videos, OmniSense introduces a lightweight spherical region of interest (SRoI) prediction algorithm to prune redundant information in 360° frames. Incorporating the video content and network dynamics, it then smartly scales vision models to analyze the predicted SRoIs with optimized resource utilization. We implement a prototype of OmniSense with commodity devices and evaluate it on diverse real-world collected 360° videos. Extensive evaluation results show that compared to resource-agnostic baselines, it improves the accuracy by 19.8% – 114.6% with similar end-to-end latencies. Meanwhile, it hits 2.0× – 2.4× speedups while keeping the accuracy on par with the highest accuracy of baselines.

Recently, video/image intelligent analytics has been widely used in industrial Artificial Intelligence (AI) applications, such as defect detection, face recognition, and security monitoring. To provide better applicability and compatibility in such applications, the embedded AI models must be developed, compiled, and deployed under different development frameworks, such as cuDNN, RKNN, etc. Unfortunately, these frameworks are supported by various Graphic Processing Unit (GPU) hardware vendors, resulting in different model parameter structures and increased development costs. To address these issues, we propose LiGo, a low cost cross-platform video/image process framework, that simplifies and accelerates video intelligent processing in practical heterogeneous hardware systems. LiGo1 provides video processing pipeline, cross-platform development environments, and unified model serving structures. We demonstrate LiGo's efficiency and flexibility in model generation and deployment through its use in supporting multiple real-world commercial industrial systems.

Clinical decision-making models have been developed to support therapeutic interventions based on medical data from either a single hospital or multiple hospitals. However, models based on multihospital data require collaboration among hospitals to integrate local data, which can result in information leakage and violate patient privacy. To address this challenge, we propose a novel approach that combines federated learning (FL) with inverse reinforcement learning (IRL) to create an efficient medical decision-making support tool while preserving patient privacy. Our approach uses an IRL algorithm with differential privacy to train a neural network-based agent on local data containing clinician trajectories, which learns a private treatment policy by observing patients’ conditions. Additionally, we integrate FL into the proposed algorithm to learn a global optimal action policy collaboratively among various smart intensive care units, overcoming data limitations at each hospital. We evaluate our approach using real-world medical data and demonstrate that it achieves superior performance in a distributed manner.

Federated learning (FL) emerges as a new distributed machine learning (ML) paradigm that enables thousands of mobile devices to collaboratively train ML models using local data without compromising user privacy. However, the FL learning quality highly relies on the data contribution from the distributed mobile devices. Therefore, a well-designed incentive mechanism with effectiveness, fairness, and reciprocity is in urgent need to guarantee the stable participation of users. In this article, we propose federated auction bandit (FedAB), an incentive and client selection strategy based on a novel multiattribute reverse auction mechanism and a combinatorial multi-armed bandit (CMAB) algorithm. First, we develop a local contribution evaluation method based on importance sampling in the FL context. We then design a novel payment mechanism that is able to preserve individual rationality and incentive compatibility (truthfulness). At last, we design a UCB-based winner selection algorithm that is proven to achieve the server’s utility maximization with fairness and reciprocity. We have conducted extensive experiments on real data sets. The results demonstrate the superiority ofFedAB, with a 10%–50% improvement in total reward, final accuracy, and convergence speed compared to state-of-the-art solutions.

The increasing awareness of privacy and the adoption of data regulations challenge the traditional trajectory publication framework in which a trusted server has access to the raw data from mobile clients. In the new untrusted environment, the clients call for much stronger data privacy preservation locally without sharing their raw data. Based on the emerging paradigm of federated analytics, we propose a Federated Analytics-based Secure Trajectory PUBlication (FASTPub) mechanism to operate in such untrusted environments. Compared with existing local differential privacy (LDP) methods, FASTPub guarantees LDP and loss-bounded$k$-anonymity simultaneously with greatly improved data utility. Specifically, FASTPub works interactively between the server and clients and iteratively builds up the trajectory without exposing raw data. Sampled clients only respond to selected trajectory fragments with randomized answers to preserve privacy as much as possible. The server then intelligently aggregates these randomized responses leveraging the intrinsic Apriori property and a Markov independent assumption of trajectory data to guide further iterations. Extensive experiments on synthetic and real-world datasets on two downstream tasks demonstrate that FASTPub gains a remarkably improved data utility compared to the existing state-of-the-art solutions.

Edge computing is promising for connected vehicles. As vehicles move, their resource demands for edge servers vary. Thus, it is necessary to reserve edge servers dynamically to meet variable demands. Existing schemes of edge-server reservation usually rely on statistical information of resource demands to make reservations; they are infeasible for connected vehicles, since such schemes are not adaptive to time-varying demands. To this end, a spatio-temporal reinforcement learning scheme called DeepReserve is developed to learn variable demands and then conduct edge-server reservation. Its design is based on the deep deterministic policy gradient algorithm of deep reinforcement learning (DRL), but is featured with several enhancements. First, the fully-connected neural network in DRL is replaced by a convolutional LSTM (ConvLSTM) network to extract spatio-temporal features of resource demands, which highly improves the prediction accuracy of resource demands. Thus, the actions in DRL (i.e., reservation decisions) can adapt to future demands. Second, an action amender is designed to ensure the actions selected by the neural network follow the spatio-temporal correlation. Finally, a training method called DR-Train is designed to stabilize the training procedure for different traffic patterns. DeepReserve is evaluated through extensive experiments on real-world datasets. Results show that it outperforms state-of-the-art approaches.

Local differential privacy (LDP) is a prominent approach and widely adopted in federated learning (FL) to preserve the privacy of local training data. It also nicely provides a rigorous privacy guarantee with computational efficiency in theory. However, a strong privacy guarantee with local differential privacy can degrade the adversarial robustness of the learned global model. To date, very few studies focus on the interplay between LDP and the adversarial robustness of federated learning. In this paper, we observe that LDP adds random noise to the data to achieve privacy guarantee of local data, and thus introduces uncertainty to the training dataset of federated learning. This leads to decreased robustness. To solve this robustness problem caused by uncertainty, we propose to leverage the promising distributionally robust optimization (DRO) modeling approach. Specifically, we first formulate a distributionally robust and private federated learning problem (DRPri). While our formulation successfully captures the uncertainty generated by the LDP, we show that it is not easily tractable. We thus transform our DRPri problem to another equivalent problem, under the Wasserstein distance-based uncertainty set, which is named the DRPri-W problem. We then design a robust and private federated learning algorithm, RPFL, to solve the DRPri-W problem. We analyze RPFL and theoretically show it satisfies differential privacy with a robustness guarantee. We evaluate algorithm RPFL by training classifiers on real-world datasets under a set of well-known attacks. Our experimental results show our algorithm RPFL can significantly improve the robustness of the trained global model under differentially private data by up to 4.33 times.

Frequent pattern mining is an important class of knowledge discovery problems. It aims at finding out high-frequency items or structures (e.g., itemset, sequence) in a database, and plays an essential role in deriving other interesting patterns, like association rules. The traditional approach of gathering data to a central server and analyze is no longer viable due to the increasing awareness of user privacy and newly established laws on data protection. Previous privacy-preserving frequent pattern mining approaches only target a particular problem with great utility loss when handling complex structures. In this paper, we take the first initiative to propose a unified federated analytics framework (FedFPM) for a variety of frequent pattern mining problems, including item, itemset, and sequence mining. FedFPM achieves high data utility and guarantees local differential privacy without uploading raw data. Specifically, FedFPM adopts an interactive query-response approach between clients and a server. The server meticulously employs the Apriori property and the Hoeffding’s inequality to generates informed queries. The clients randomize their responses in the reduced space to realize local differential privacy. Experiments on three different frequent pattern mining tasks demonstrate that FedFPM achieves better performances than the state-of-the-art specialized benchmarks, with a much smaller computation overhead.

With the advancement of deep learning techniques, major cloud providers and niche machine learning service providers start to offer their cloud-based machine learning tools, also known as machine learning as a service (MLaaS), to the public. According to our measurement, for the same task, these MLaaSes from different providers have varying performance due to the proprietary datasets, models, etc. Federating different MLaaSes together allows us to improve the analytic performance further. However, naively aggregating results from different MLaaSes not only incurs significant momentary cost but also may lead to sub-optimal performance gain due to the introduction of possible false-positive results. In this paper, we propose Armol, a framework to federate the right selection of MLaaS providers to achieve the best possible analytic performance. We first design a word grouping algorithm to unify the output labels across different providers. We then present a deep combinatorial reinforcement learning based-approach to maximize the accuracy while minimizing the cost. The predictions from the selected providers are then aggregated together using carefully chosen ensemble strategies. The real-world trace-driven evaluation further demonstrates that Armol is able to achieve the same accuracy results with 67% less inference cost.

Node embedding aims to map nodes in the complex graph into low-dimensional representations. The real-world large-scale graphs and difficulties of labeling motivate wide studies of unsupervised node embedding problems. Nevertheless, previous effort mostly operates in a centralized setting where a complete graph is given. With the growing awareness of data privacy, data holders who can be represented by one vertex in the graph and are only its neighbors demand greater privacy protection. In this paper, we introduce FedWalk, a random-walk-based unsupervised node embedding algorithm that operates in such a node-level visibility graph with raw graph information remaining locally. FedWalk is designed to offer centralized competitive graph representation capability with data privacy protection and great communication efficiency. FedWalk instantiates the prevalent federated paradigm and contains three modules. We first design a hierarchical clustering tree (HCT) constructor to extract the structural feature of each node. A dynamic time warping algorithm seamlessly handles the structural heterogeneity across different nodes. Based on the constructed HCT, we then design a random walk generator, wherein a sequence encoder is designed to preserve privacy and a two-hop neighbor predictor is designed to save communication cost. The generated random walks are then used to update node embedding based on a SkipGram model. Extensive experiments on two large graphs demonstrate that FedWalk achieves competitive representativeness as a centralized node embedding algorithm does with only up to 1.8% Micro-F1 score and 4.4% Marco-F1 score loss while reducing about 6.7 times of inter-device communication per walk.

Over the past decades, 360-degree videos have attracted wide interest for the immersive experience they bring to viewers. The rising of high-resolution 360-degree videos greatly challenges the traditional video streaming systems in limited network environments. Given the limited bandwidth, tile-based video streaming with adaptive bitrate selection has been widely studied to improve the Quality of Experience (QoE) of viewers by tiling the video frames and allocating different bitrates for tiles inside and outside viewers' viewports. Existing solutions for viewport prediction and bitrate selection train general models without catering to the intrinsic need for personalization. In this paper, we present the first meta-learning-based personalized 360-degree video streaming framework. The commonality among viewers of different viewing patterns and QoE preferences is captured by efficient meta-network designs. Specifically, we design a meta-based long-short term memory model for viewport prediction and a meta-based reinforcement learning model for bitrate selection. Extensive experiments on real-world datasets demonstrate that our framework not only outperforms the state-of-the-art data-driven approaches in prediction accuracy by 11% on average and improves QoE by 27% on average, but also quickly adapts to users with new preferences with on average 67%-88% less training epochs.

Conventional model quantization methods use a fixed quantization scheme to different data samples, which ignores the inherent"recognition difficulty" differences between various samples. We propose to feed different data samples with varying quantization schemes to achieve a data-dependent dynamic inference, at a fine-grained layer level. However, enabling this adaptive inference with changeable layer-wise quantization schemes is challenging because the combination of bit-widths and layers is growing exponentially, making it extremely difficult to train a single model in such a vast searching space and use it in practice. To solve this problem, we present the Arbitrary Bit-width Network (ABN), where the bit-widths of a single deep network can change at runtime for different data samples, with a layer-wise granularity. Specifically, first we build a weight-shared layer-wise quantizable "super-network" in which each layer can be allocated with multiple bit-widths and thus quantized differently on demand. The super-network provides a considerably large number of combinations of bit-widths and layers, each of which can be used during inference without retraining or storing myriad models. Second, based on the well-trained super-network, each layer's runtime bit-width selection decision is modeled as a Markov Decision Process (MDP) and solved by an adaptive inference strategy accordingly. Experiments show that the super-network can be built without accuracy degradation, and the bit-widths allocation of each layer can be adjusted to deal with various inputs on the fly. On ImageNet classification, we achieve 1.1% top1 accuracy improvement while saving 36.2% BitOps.

Cross-camera video analytics is a major video analytic task that associates and analyzes information across multiple cameras. Existing studies exploit static temporal and spatial correlation among different cameras to reduce the searching space and thus accelerate searching. However, in practice, correlation among cameras changes over time, making the previous static approach inefficient. In this paper, we present the first work to support dynamic correlation estimation based on a few samples. Specifically, we formulate the correlation estimation problem as a mean-field game and interpret the cross-camera correlation as a novel mean-field term to approximate the target cameras' density field. Despite the significant complexity to solve the coupled partial differential equations in the game, we propose a G-prox primal-dual hybrid gradient algorithm to solve it efficiently. Our derived correlation models can guide searching with controllable granularity and solely rely on the knowledge of the initial correlation and destination correlation information. Extensive experiments on a real-world dataset demonstrate that, with the help of our dynamic correlation models, the overall workload can be reduced by 36% in general.

5G has seen rapid growth recently, attracting several measurement studies on its coverage, connectivity and quality of service. However, there is still a lack of understanding of 5G's capabilities and potential impacts from a content provider (CP)'s perspective. This paper fills in this gap by studying 5G networks used by over 23 million users in one year in Kuaishou, a popular crowdsourced live streaming platform. Our measurements provide the following discoveries. i) Standalone (SA) 5G generally provides end-to-end performance improvement as compared with 4G or non-SA (NSA) 5G, but its advantage depends on both the number of cellular users and CP-level configurations. ii) In the radio access network, SA 5G is more sensitive to access density but has better handover tolerance. iii) Controlled experiments with 29 mobile device models on energy consumption refute some "conventional wisdom," including that 5G always consumes more power. iv) Traceroute-based active experiments in over 300 cities show that although users are "closer" to the internet in SA 5G, their end-to-end latency may not benefit from that. Furthermore, we show new design space for 5G participants and provide a 5G-aware rebuffer strategy tested by 9 million viewers in Kuaishou, with a 7% reduction in rebuffer proportion.

The local model poisoning attack is an attack to manipulate the shared local models during the process of distributed learning. Existing defense methods are passive in the sense that they try to mitigate the negative impact of the poisoned local models instead of eliminating them. In this paper, we leverage the new federated analytics paradigm, to develop a proactive defense method. More specifically, federated analytics is to collectively carry out analytics tasks without disclosing local data of the edge devices. We propose a Federated Anomaly Analytics enhanced Distributed Learning (FAA-DL) framework, where the clients and the server collaboratively analyze the anomalies. FAA-DL firstly detects all the uploaded local models and splits out the potential malicious ones. Then, it verifies each potential malicious local model with functional encryption. Finally, it removes the verified anomalies and aggregates the remaining to produce the global model. We analyze the FAA-DL framework and show that it is accurate, robust, and efficient. We evaluate FAA-DL by training classifiers on MNIST and Fashion-MNIST under various local model poisoning attacks. Our experiment results show FAA-DL improves the accuracy of the learned global model under strong attacks up to 6.90 times and outperforms the state-of-the-art defense methods with a robustness guarantee.

Considering network slicing in a cellular network, one of the most intriguing tasks is slice enforcement over air interfaces across multiple cells. The challenges lie in several aspects. First, resources allocated to different slices must achieve soft isolation at the link level. Second, users’ diverse QoS requirements must be satisfied even when communication links experience fading and interference. Third, long-term slicing policies must be conformed, no matter how unbalanced they are. To address these challenges, link-level slice enforcement is first formulated as a resource allocation problem that minimizes radio resource consumption while ensuring link-level soft slice isolation, guaranteeing users’ diverse QoS requirements, and conforming to slicing policies. Next, this problem is tackled via a deep reinforcement learning (DRL) based approach, through which LinkSlice is designed as an iterative two-stage algorithm. The first stage determines transmission rates for each link based on DRL. It is embedded with a graph neural network (GNN) to characterize link interference. Based on the transmission rates from the first stage, the second stage allocates resources to each slice. Performance results show that LinkSlice converges quickly to a near-optimal solution. It gracefully tackles the three challenges of link-level slice enforcement while further improving throughput by 18.5%.

Video processing plays an essential role in a wide range of cloud-based applications. It typically involves multiple pipelined stages, which well fits the latest fine-grained serverless computing paradigm if properly configured to match the cost and delay constraints of video. Existing configuration tools, however, are primarily developed for traditional virtual machine clusters with general workloads. This paper presents CharmSeeker, an automated configuration tuning tool for serverless video processing pipelines. We first carefully examine the key steps and the performance bottlenecks for video processing over modern serverless platforms. Then, we identify the configuration space for processing pipelines and leverage a carefully designed Sequential Bayesian Optimization search scheme to identify promising configurations. We further address the practical challenges toward integrating our solution into real-world systems and develop a prototype with AWS Lambda. Evaluation results show that CharmSeeker can find out the optimal or near-optimal configurations that improve the relative processing time up to 408.77%. It is also more robust and scalable to various video processing pipelines compared with state-of-the-art solutions.

Edge computing is promising to provide computational resources for connected vehicles. Resource demands for edge servers vary due to vehicle mobility. It is then challenging to reserve edge servers to meet variable demands. Existing schemes rely on statistical information of resource demands to determine edge server reservation. They are infeasible in practice, since the reservation based on statistics cannot adapt to time-varying demands. In this paper, a spatio-temporal reinforcement learning scheme called DeepReserve is developed to learn variable demands and then reserve edge servers accordingly. DeepReserve is adapted from the deep deterministic policy gradient algorithm with two major enhancements. First, by observing that the spatio-temporal correlation in vehicle traffic leads to the same property in resource demands of CVs, a convolutional LSTM network is employed to encode resource demands observed by edge servers for inference of future demands. Second, an action amender is designed to make sure an action does not violate spatio-temporal correlation. We also design a new training method, i.e., DR-Train, to stabilize the training procedure. DeepReserve is evaluated via experiments based on real-world datasets. Results show it achieves better performance than state-of-the-art approaches that require accurate demand information.

The emerging federated optimization paradigm performs data mining or artificial intelligence techniques locally on the edge devices, enabling scientists and engineers to utilize the blooming edge data with privacy protection. In such a paradigm, since data cannot be shared or gathered, data heterogeneity naturally emerges, which significantly degrades the performance of federated optimization, ultimately leading to poor quality of federated services. In this paper, we present the first work on characterizing the data heterogeneity in the framework of federated analytics, i.e., to collectively carry out analytics tasks without raw data sharing, and use the information to create a desirable data environment via intelligent client selection. Our proposed Analytics-driven Client Selection framework, named FedACS, tackles the data heterogeneity problem in three steps. First, clients are in charge of generating insights about local data without disclosure of sensitive information. Then, the server uses these insights to infer the situation of clients’ data heterogeneity based on the Hoeffding’s inequality. Finally, a dueling bandit is formulated to intelligently select clients with slighter data heterogeneity to form a desirable client pool. FedACS can be universally applied to all kinds of federated optimization tasks, and gains benefits including privacy protection, infrastructure reuse, and client load reduction. To test its efficiency, we further customize it to assist federated learning, a popular scenario of federated optimization. According to experiment results, FedACS reduces the accuracy degrading by up to 65.6%, and speeds up the convergence for up to 2.4 times.

The ever-growing deployment scale of surveillance cameras and the users' increasing appetite for real-time queries have urged online video analytics. Synergizing the virtually unlimited cloud resources with agile edge processing would deliver an ideal online video analytics system; yet, given the complex interaction and dependency within and across video query pipelines, it is easier said than done. This paper starts with a measurement study to acquire a deep understanding of video query pipelines on real-world camera streams. We identify the potentials and practical challenges towards cloud-edge collaborative video analytics. We then argue that the newly emerged serverless computing paradigm is the key to achieve fine-grained resource partitioning with minimum dependency. We accordingly propose CEVAS, a Cloud-Edge collaborative Video Analytics system empowered by fine-grained Serverless pipelines. It builds flexible serverless-based infrastructures to facilitate fine-grained and adaptive partitioning of cloud-edge workloads for multiple concurrent query pipelines. With the optimized design of individual modules and their integration, CEVAS achieves real-time responses to highly dynamic input workloads. We have developed a prototype of CEVAS over Amazon Web Services (AWS) and conducted extensive experiments with real-world video streams and queries. The results show that by judiciously coordinating the fine-grained serverless resources in the cloud and at the edge, CEVAS reduces 86.9% cloud expenditure and 74.4% data transfer overhead of a pure cloud scheme and improves the analysis throughput of a pure edge scheme by up to 20.6%. Thanks to the fine-grained video content-aware forecasting, CEVAS is also more adaptive than the state-of-the-art cloud-edge collaborative scheme.

We are now in an information era and the volume of data is growing explosively. However, due to privacy issues, it is very common that data cannot be freely shared among the data generating.Federated analyticswas recently proposed aiming at deriving analytical insights among data-generating devices without exposing the raw data, but the intermediate analytics results. Note that the computing resources at the data generating devices are limited, thus making on-device execution of computing-intensive tasks challenging. We thus propose to apply the digital twin technique, which emulates the resource-limited physical/end side, while utilizing the rich resource at the virtual/computing side. Nevertheless, how to use the digital twin technique to assist federated analytics while preserving distributed data privacy is challenging. To address such a challenge, this work first formulates a problem on digital twin-assisted federated distribution discovery. Then, we propose a federated Markov chain Monte Carlo with a delayed rejection (FMCMC-DR) method to estimate the representative parameters of the global distribution. We combine a rejection–acceptance sampling technique and a delayed rejection technique, allowing our method to be able to explore the full state space. Finally, we evaluate FMCMC-DR against the Metropolis–Hastings (MH) algorithm and random walk Markov chain Monte Carlo method (RW-MCMC) using numerical experiments. The results show our algorithm outperforms the other two methods by 50% and 95% contour accuracy, respectively, and has a better convergence rate.

The explosion of online shopping brings great challenges to traditional logistics industry, where the massive parcels and tight delivery deadline impose a large cost on the delivery process, in particular the last mile parcel delivery. On the other hand, modern cities never lack transportation resources such as the private car trips. Motivated by these observations, we propose a novel and effective last mile parcel delivery mechanism through car trip sharing, to leverage the available private car trips to incidentally deliver parcels during their original trips. To achieve this, the major challenges lie in how to accurately estimate the parcel delivery trip cost and assign proper tasks to suitable car trips to maximize the overall performance. To this end, we develop Car4Pac, an intelligent last mile parcel delivery system to address these challenges. Leveraging the real-world massive car trip trajectories, we first build up a 3D (time-dependent, driver-dependent and vehicle-dependent) landmark graph that accurately predicts the travel time and fuel consumption of each road segment. Our prediction method considers not only traffic conditions of different times, but also driving skills of different people and fuel efficiencies of different vehicles. We then develop a two-stage solution towards the parcel delivery task assignment, which is optimal for one-to-one assignment and yields high-quality results for many-to-one assignment. Our extensive real-world trace driven evaluations further demonstrate the superiority of our Car4Pac solution.

Today’s anywhere and anytime broadband connection and audio/video capture have boosted the deployment of crowdsourced livecast services (orcrowdcast). Bridging a massive amount of geo-distributed broadcasters and their fellow viewers, such representatives as Twitch.tv, Youtube Gaming, and Inke.tv, have greatly changed the generation and distribution landscape of streaming content. They also enable rich online interactions among the crowd, and strive to offer personalized Quality-of-Experience (QoE) for individual viewers. Given the ultra-large scale and the dynamics of the crowd, personalizing QoE however is much more challenging than in early generation streaming services. The rich interactions among the broadcasters, viewers, and the network system, on the other hand, also offer invaluable data that could be utilized towards informed management. This paper presentsDeepCast, an edge-assisted crowdcast framework that explores the sheer amount of viewing data towards intelligent decisions for personalized QoE demands. DeepCast seamlessly integrates cloud, CDN, and edge servers for crowdcast content distribution, and advocates a data-driven design that extracts the hidden information from the complex interactions among the system components. Through deep reinforcement learning (DRL), it automatically identifies the most suitable strategies for viewer assignment and transcoding at edges. We collect multiple real-world datasets and evaluate the performance of DeepCast with trace-driven experiments. The results demonstrate its flexibility and effectiveness towards better personalized QoE and lower cost for crowdcast systems.

Recent years have seen booming development and great success in interactive crowdsourced livecast (i.e., crowdcast). Different from traditional livecast services, crowdcast is featured with tremendous video contents at the broadcaster side, highly diverse viewer side content watching environments/preferences as well as viewers' personalized quality of experience (QoE) demands (e.g., individual preferences for streaming delays, channel switching latencies and bitrates). This imposes unprecedented key challenges on how to flexibly and cost-effectively accommodate the heterogeneous and personalized QoE demands for the mass of viewers. In this paper, we propose DeepCast, an edge-assisted crowdcast framework, which makes intelligent decisions at edges based on the massive amount of real-time information from the network and viewers to accommodate personalized QoE with minimized system cost. Given the excessive computation complexity in this context, we propose a data-driven deep reinforcement learning (DRL) based solution that can automatically learn the best suitable strategies for viewer scheduling and transcoding selection. To our best knowledge, DeepCast is the first edge-assisted framework that applies the advance of DRL to explicitly accommodate personalized QoE optimization for crowdcast services. We collect multiple real-world datasets and evaluate the performance of DeepCast using trace-driven experiments. The results demonstrate the superiority of our DeepCast framework and its DRL-based solution.

The growing demand for video processing and the advantages in scalability and cost reduction brought by the emerging serverless computing have attracted significant attention in serverless computing powered video processing. However, how to implement and configure serverless functions to optimize the performance and cost of video processing applications remains unclear. In this paper, we explore the configuration and implementation schemes of typical video processing functions deployed to the serverless platforms and quantify their influence on the execution duration and monetary cost from a developer's perspective. Our measurement reveals that memory configuration is non-trivial. Dynamic profiling of workloads is necessary to find the best memory configuration. Moreover, compared with calling external video processing APIs, implementing these services locally in serverless functions can be competitive. We also find that the performance of video processing applications could be affected by the underlying infrastructure. Our work provides guidelines for further function-level optimization and complements the existing measurement studies for both serverless computing and video processing.

Deep learning has been applied in many recent advanced applications in the field of transportation, finance and medicine. These applications require significant computation resources and large-scale training samples. Cloud becomes a natural choice for conducting these learning tasks due to its abundant resources. However, deeper penetration of deep learning techniques in mission critical applications, like driverless car, calls for stricter time requirement to guarantee its interaction and larger amount of dataset for training to guarantee its accuracy, which cannot be easily satisfied by the cloud and makes the network transmission become the bottleneck. Edge learning emerges to be a promising direction to reduce data transmission time by processing and compressing the raw data at the edge of the network, while brings the concern of accuracy reduction at the meantime. To balance this tradeoff under cloud-edge architecture, we study a task scheduling problem for reducing weighted transmission time which takes learning accuracy into consideration. We also propose efficient scheduling algorithms which are able to achieve up to 50% reduction in makespan with extensive trace-driven simulations.

Nowadays online shopping has become explosively popular and the vast numbers of generated packages have brought great challenges to the traditional logistics industry, especially the last mile package delivery. Traditional delivery approaches rely on dedicated couriers for package dispatch, while the labor cost is quite expensive and the quality is hard to guarantee due to the diverse delivery addresses and tight deadlines. On the other hand, modern cities are full of available transportation resources such as private car trips. The mobile crowdsourcing through 4G/5G and vehicle-related communications enables the vehicle resources to be connected as an intelligent transportation system. As such, we believe ridesharing will be a core service for connected vehicles, which we refer to as Ridesharing as a Service (RaaS). In this paper, we focus on the quality of service (QoS) of RaaS in the last mile package delivery. Mining from real-world car trips, we build up a citywide routing graph and conduct a personalized travel cost prediction considering both the travel time of each driver and the fuel consumption of each vehicle. We then design an online algorithm to assign proper package delivery tasks to the submitted car trips, aiming to maximize the utility of the ridesharing service provider. Our extensive real-world trace-driven evaluations further demonstrate the superiority of our RaaS based package delivery.

Despite that IaaS users are busy scaling up/out their cloud instances to meet the ever-increasing demands, the dynamics of their demands, as well as the coarse-grained billing options offered by leading cloud providers, have led to substantial instance underutilization in both temporal and spatial domains. This paper theoretically examines an instance subletting service, where underutilized instances are leased to others within user-specified periods. Serving as a secondary market that complements the existing instance market of IaaS providers,we specifically identify the theoretical challenges in instance subletting services, and design an online auction mechanism tomake allocation and pricing decisions for the instances to besublet. Our mechanism guarantees truthfulness and individualrationality with the best possible competitive ratio. Extensivetrace-driven simulations show that our proposed mechanismachieves significant performance gains in both cost and socialwelfare.

In the public cloud market, there has been a constant battle over the billing options of the cloud instances between their providers and their users. The users generally have to pay for the entire billing cycle even on fractional usage. Ideally, the residual life-cycles should be resalable by the users, which demands efficient resource consolidation and multiplexing; otherwise, the revenue and use cases are confined by the transient nature of the instances. This paper presents HARV, a novel cloud service that facilitates the management and trade of cloud instances through a third-party platform to run buyers' tasks. The platform relies on hybrid virtualization, an infrastructure layout integrating both the hypervisor-based virtualization and lightweight containerization. It further incorporates a truthful online auction mechanism for instance trading and resource allocation. Our design achieves efficient resource consolidation with no need for provider-level support, and we have deployed a prototype of HARV on the Amazon EC2 public cloud. Our evaluations on both micro-benchmarks and real-life workloads reveal that applications experience negligible performance overhead when hosted on HARV. Trace-driven simulations further show that HARV can achieve substantial cost savings.

Ever-increasing research effort has been dedicated to multi-group radio frequency identification (RFID) systems where all tags are partitioned into multiple groups, such as group-level queries, and multi-group missing tag detection. However, it is assumed in the existing work that all tags know their individual group IDs, which thus leaves group labeling problem unaddressed. To tackle the under-investigated problem, this paper is devoted to devising an efficient group labeling protocol to inform each tag of its corresponding group ID fast and accurately. To this end, we employ multiple seeds to build a Composite Indicator Vector (CIV) indicating the assigned seed in each slot, which reduces transmissions of useless information and thus improves time efficiency. Specifically, we first theoretically show that the Seed Assignment Problem (SAP) arising in establishing the CIV is NP-hard and then develop a myopic approximation algorithm. Finally, the simulation results confirm the superiority of the proposed protocol over the state-of-the-art solution in terms of time efficiency.

During the past decade, we have witnessed the pervasive penetration of mobile smart devices such as smartphones, tablets, and wearable devices, which significantly enrich Internet applications and improve user experience. In the foreseeable future, mobile smart devices are predicted to take up over 50% of global devices/connections and surpass 4/5 of mobile data traffic by 2021 [1]. Such mobile smart devices as smartphones, phablets, and tablets, undoubtedly reshape the way that users access Internet services, e.g., web browsing. Different from traditional desktop applications, in which users interact via interfaces like large displays, keyboards, and mouses, mobile applications require users to enter the inputs through touch screens and allow them to view the outputs on limited size of displays. This distinct feature introduced by mobile hardware interfaces brings both challenges and opportunities to mobile-based Internet applications. On one hand, mobile service providers should prepare multiple copies of media contents with different resolutions and even multiple versions of application UI layouts to fit various sizes of screens on heterogeneous devices. On the other hand, as media contents are usually organized in certain order in mobile-based Internet applications, it is possible to predict the viewing region (referred as viewport hereafter) given the user inputs and the fixed size of display.

There has been explosive growth in Mobile Personal Livestreaming (MPL) market since 2016. MPL services are booming not only because they introduce the popular live content by spontaneous and personalized broadcasters, but also because they are deliberately designed to be the innovative social networking service (SNS) platforms. The latter is a very important aspect that distinguishes MPL from the traditional livestreaming services. In this paper, we study the social networking of a large scale MPL service “Inke” (with more than 200 million registered users, 15 million daily active users) in China. By analyzing the dataset we crawl and the features of Inke app, we show that the social media stickiness of Inke comes from three aspects: the follower-followee model, the virtual-gift-based incentive mechanism, and the multi-perspective interactivity between broadcasters and viewers. First, Inke introduces the follower-followee model rather than the traditional broadcaster-viewer model, and every user in Inke can be a broadcaster. This makes MPL have some different patterns from both the traditional livestreaming services and SNS platforms. Second, Inke use virtual gift giving and user ranking as its incentive mechanism. Our measurement results show that this mechanism can indeed enhance user stickiness. Furthermore, Inke incorporates a variety of features during broadcasting to strengthen interactivity. The insight we gain in this paper has important implications for both existing and future designs.

Live video contents in crowdsourced live streaming services are transcoded into multiple quality versions to better service viewers with different network and device configurations. Cloud computing becomes a natural choice to handle this transcoding service due to its elasticity and significant computational power. However, given the huge concurrent channel numbers in this crowdsourced live streaming service, even the cloud becomes significantly expensive for providing transcoding services to the whole community. In this poster, after observing that abundant computational resources reside in end viewers, we propose a Cloud-Crowd collaborative system, C2, which incentivizes idle end-viewers to join with the cloud to do video transcoding. Specifically, we propose an auction mechanism to carefully select stable viewers and determine the proper payment for them. Desirable economic properties, like incentive compatibility, can be achieved in our mechanism. Large-scale trace-driven simulations further demonstrate the superiority of our mechanisms in cost reduction and service stability.

In the emerging crowd sourced live cast services, numerous amateur broadcasters live stream their video contents to worldwide viewers and constantly interact with them through chat messages. Live video contents are transcoded into multiple quality versions to better service viewers with different network and device configurations. Cloud computing becomes a natural choice to handle these computational intensive tasks due to its elasticity and the "pay-as-you-go" billing model. However, given the significantly large number of concurrent channel numbers and the diverse viewer geo-distributions in this new crowd sourced live cast service, even the cloud becomes significantly expensive to cover the whole community and inadequate in fulfilling the latency requirement. In this paper, after observing the abundant computational resources residing in end viewers, we propose a Cloud-Crowd collaborative system, C2, which combines end viewers with cloud to perform video transcoding in a cost-efficient way. To quantify the heterogeneity and uncertainty of viewers and pass the asymmetric information barrier, we incorporate statistical descriptions into our bidding language and design truthful auctions to recruit stable viewers with appropriate incentives. We further tailor redundancy strategies for workloads with different Quality of Service requirements to improve the stability of our system. Desirable economic properties, like social efficiency, ex-post incentive compatibility, individual rationality, are proved to be guaranteed in our studied scenarios. Using traces captured from the popular Twitch platform, we show that C2 achieves up to 93% more cost saving than a pure cloud-based solution, and significantly outperforms other baseline approaches in both social welfare and system stability.

Even with modern physical-layer technologies in LTE networks, the capacity of cellular networks is still far from sufficient to satisfy the insatiable bandwidth demand of mobile applications. Owing to common interests among mobile users, Device-to-Device (D2D) communication has emerged as a viable alternative to offload cellular traffic, with the promise of substantially alleviating the need for cellular network bandwidth. In this paper, we first carry out an extensive theoretical analysis based on a game theoretic approach, and show that the objective of maximized cellular offloading is equivalent to maximizing the social welfare in a trading network, where the content to be shared is the commodity, and mobile users are buyers or sellers. We next design Rally, a set of distributed strategies that can converge to a sub game perfect Nash equilibrium in the content sharing game. Both our theoretical analyses and simulation results have shown the effectiveness of Rally, in that it can indeed maximize cellular traffic offloading through D2D communication.

Despite the growing deployment of 4G networks, the capacity of cellular networks is still insufficient to satisfy the ever-increasing bandwidth demand of mobile applications. Given the common interest of mobile users, Device-to-Device (D2D) communication has emerged as a promising solution to offload cellular traffic and enable proximity-based services. One of the main detriments for D2D communication is the lack of incentive for mobile users to share their content, since such sharing inevitably consumes limited resources and potentially jeopardizes user privacy. In this paper, we study the incentive problem in D2D communications. Specifically, we model the incentive in offloading scenario as an auction game. A trading network is constructed between an eNB and users, in which auctions are conducted to group offloading users and determine proper rewards. We further design a randomized auction mechanism to guarantee system efficiency and truthfulness. Extensive experiments verify the effectiveness of our mechanism in that it achieves a significant performance gain in comparison with baseline methods.