Extended Reality (XR) experiences involve interactions between users, the real world, and virtual content. A key step to enable these experiences is the XR headset sensing and estimating the user's pose in order to accurately place and render virtual content in the real world. XR headsets use multiple sensors (e.g., cameras, inertial measurement unit) to perform pose estimation and improve its robustness, but this provides an attack surface for adversaries to interfere with the pose estimation process. In this paper, we create and study the effects of acoustic attacks that create false signals in the inertial measurement unit (IMU) on XR headsets, leading to adverse downstream effects on XR applications. We generate resonant acoustic signals on a HoloLens 2 and measure the resulting perturbations in the IMU readings, and also demonstrate both finegrained and coarse attacks on the ORB-SLAM3 and an open-source XR system (ILLIXR). With the knowledge gleaned from attacking these open-source frameworks, we demonstrate four end-to-end proof-of-concept attacks on a HoloLens 2: manipulating user input, clickjacking, zone invasion, and denial of user interaction. Our experiments show that current commercial XR headsets are susceptible to acoustic attacks, raising concerns for their security.

In this work, we present L3GS, a framework for efficient 3D scene delivery with 3D Gaussian splats [5]. While 3D Gaussian splats achieve a balance between visual fidelity and rendering efficiency, their massive data size still limits realtime deployment. L3GS addresses this by leveraging scene layering, predictive viewport and bandwidth estimation, and priority-based scheduling to progressively deliver the most essential splats first. This demo will showcase L3GS's 3D scene delivery performance under controlled network conditions and compare it with various baselines.

iGYM is an augmented reality exercise game for inclusive play that allows people with and without wheelchairs to participate equally in a soccer game with a projected virtual field and ball. However, currently iGYM requires all players to be co-located and lacks capabilities for remote play. In this work, we describe a networked iGYM implementation that allows teams of players to play with each other remotely. The key networking challenge is meeting tight end-to-end latency requirements for interactive play over the Internet. We demonstrate a portable tabletop version of iGYM, implemented in Unity and ROS2, using a distributed authority model to keep track of ownership and propagate state updates about players, game objects, and scores. Each client receives local player tracking updates and spawns player peripersonal circles under its own authority; a shared ball object is synchronized under server ownership with clientside interpolation. The demo GUI will let attendees inject artificial network delay to explore its impact on interactivity.

Traditional 3D content representations include dense point clouds that consume large amounts of data and hence network bandwidth, while newer representations such as neural radiance fields suffer from poor frame rates due to their nonstandard volumetric rendering pipeline. 3D Gaussian splats (3DGS) can be seen as a generalization of point clouds that meet the best of both worlds, with high visual quality and efficient rendering for real-time frame rates. However, delivering 3DGS scenes from a hosting server to client devices is still challenging due to high network data consumption (e.g., 1.5 GB for a single scene). The goal of this work is to create an efficient 3D content delivery framework that allows users to view high quality 3D scenes with 3DGS as the underlying data representation. The main contributions of the paper are: (1) Creating new layered 3DGS scenes for efficient delivery, (2) Scheduling algorithms to choose what splats to download at what time, and (3) Trace-driven experiments from users wearing virtual reality headsets to evaluate the visual quality and latency. Our system for Layered 3D Gaussian Splats delivery (L3GS) demonstrates high visual quality, achieving 16.9% higher average SSIM compared to baselines, and also works with other compressed 3DGS representations. The code is available at https://github.com/mavenslab/layered_3d_gaussian_splats.

Small language models (SLMs) struggle to learn complex reasoning behaviors, especially when high-quality traces are scarce or difficult to learn from. A typical approach for training such models combines a supervised fine-tuning (SFT) stage, often to distill reasoning capabilities from a larger model, followed by a reinforcement learning (RL) stage such as Group Relative Policy Optimization (GRPO). In this paper, we investigate the fundamental limitations of this SFT + RL paradigm and propose methods to overcome them. Using a toy student-expert model over Markov chains, we demonstrate that the SFT + RL strategy can fail completely when (1) the expert's traces are too difficult for the small model to express, or (2) the small model's initialization achieves exponentially sparse rewards as task complexity grows. To address these, we introduce BREAD, a GRPO variant that bridges SFT and RL via partial expert guidance and branch rollouts. When self-generated traces fail, BREAD adaptively inserts short expert prefixes/hints, allowing the small model to complete the rest of the reasoning path, and ensuring that each update includes at least one successful trace. This mechanism both densifies the reward signal and induces a natural learning curriculum. BREAD requires fewer than 40\% of ground-truth traces, consistently outperforming standard GRPO while speeding up the training by about 3$\times$. Importantly, we find that BREAD helps the model solve problems that are otherwise unsolvable by the SFT + RL strategy, highlighting how branch rollouts and expert guidance can aid SLM reasoning.

Crowd monitoring in sports stadiums is important to enhance public safety and improve audience experience. Existing approaches mainly rely on cameras and microphones, which can cause significant disturbances and often raise privacy concerns. In this paper, we sense floor vibration, which provides a less disruptive and more non-intrusive way of crowd sensing, to predict crowd behavior. However, since the vibration-based crowd monitoring approach is newly developed, one main challenge is the lack of training data due to sports stadiums are usually large public spaces with complex physical activities.

Evaluation lkthrough with Privacy Experts manner?How can we optimize negotiations of AR sensing capabilities to balance multiple usersʼ UX and privacy needs in a fine-grained manner?How could we mediate negotiations of AR sensing capabilities to balance multiple individualsʼ UX and privacy needs in a fine-grained manner?Figure 1: Augmented reality glasses pose privacy risks for co-located individuals, but today, their use in public spaces is governed solely by the wearer.Our work explores how to facilitate multi-user negotiations of AR sensing capabilities, formulating this process as an optimization approach to maintain core AR functionality while achieving a balance, or Equilibrium, with privacy.

Video streaming applications have become immensely popular, leading to increasing user expectations for high-quality services. Extensive work has been conducted in the areas of Quality of Experience (QoE) modeling and Adaptive Bitrate (ABR) algorithms to meet this demand. While learningbased approaches have demonstrated substantial progress using large-scale datasets, existing QoE models often focus on systemlevel metrics such as bitrate and resolution within the playback buffer, neglecting the quality as perceived by the human eye. Simultaneously, many learning-based ABR algorithms exhibit limited robustness in dynamic environments due to their reliance on a one-size-fits-all strategy, which fails to adapt effectively to complex, real-world conditions. In this paper, we propose an integrated system that addresses these limitations by combining an accurate QoE model with an environment-robust adaptation algorithm to enhance user satisfaction in diverse and dynamic environments. First, we introduce RetQoE, a novel approach that accurately estimates the user’s actual QoE by focusing on the quality of video content as perceived by the viewer, rather than on conventional system metrics. Then, we design PVA, a meta-reinforcement learning-based adaptation that rapidly adjusts its policy to varying environments. We systematically integrate RetQoE and PVA, enabling PVA to update its policy with feedback from RetQoE in just a few steps online. We demonstrate the effectiveness of RetQoE+PVA through extensive evaluations in diverse environments, outperforming conventional learning-based algorithms across various metrics.

Augmented reality (AR) seamlessly overlays virtual objects onto the real world, enabling an exciting new range of applications. Multiple users view and interact with virtual objects, which are replicated and shown on each user’s display. A key requirement of AR is that the replicas should be quickly updated and converge to the same state; otherwise, users may have laggy or inconsistent views of the virtual object, which negatively affects their experience. A second key requirement is that the movements of virtual objects in space should preserve certain integrity properties either due to physical boundaries in the real world, or privacy and safety preferences of the user. For example, a virtual cup should not sink into a table, or a private virtual whiteboard should stay within an office. The challenge tackled in this paper is the coordination of virtual objects with low latency, spatial integrity properties and convergence. We introduce “well-organized” replicated data types that guarantee these two properties. Importantly, they capture a local notion of conflict that supports more concurrency and lower latency. To implement well-organized virtual objects, we introduce a credit scheme and replication protocol that further facilitate local execution, and prove the protocol’s correctness. Given an AR environment, we automatically derive conflicting actions through constraint solving, and statically instantiate the protocol to synthesize custom coordination. We evaluate our implementation, H ambazi , on off-the-shelf Android AR devices and show a latency reduction of 30.5-88.4% and a location staleness reduction of 35.6-75.6%, compared to three baselines, for varying numbers of devices, AR environments, request loads, and network conditions.

Modern classification problems exhibit heterogeneities across individual classes: Each class may have unique attributes, such as sample size, label quality, or predictability (easy vs difficult), and variable importance at test-time. Without care, these heterogeneities impede the learning process, most notably, when optimizing fairness objectives. Confirming this, under a gaussian mixture setting, we show that the optimal SVM classifier for balanced accuracy needs to be adaptive to the class attributes. This motivates us to propose CAP: An effective and general method that generates a class-specific learning strategy (e.g.~hyperparameter) based on the attributes of that class. This way, optimization process better adapts to heterogeneities. CAP leads to substantial improvements over the naive approach of assigning separate hyperparameters to each class. We instantiate CAP for loss function design and post-hoc logit adjustment, with emphasis on label-imbalanced problems. We show that CAP is competitive with prior art and its flexibility unlocks clear benefits for fairness objectives beyond balanced accuracy. Finally, we evaluate CAP on problems with label noise as well as weighted test objectives to showcase how CAP can jointly adapt to different heterogeneities.

In world-locked mixed reality (MR), virtual content is locked in place with respect to the real world. Pose estimation is a key component to create world-locked MR experiences by estimating the device's position and orientation in order to render the virtual content accordingly. Current methods of evaluating world-locked MR include user studies, which are time consuming, and absolute trajectory error (ATE), which does not directly represent what is shown on the user's display. In this work, we propose ReplayAR, a tool that can replay user movement traces and output the corresponding visualizations (renderings) of the MR display. ReplayAR can be used to compare renderings from different MR pose estimation methods side by side, using our proposed Visual Difference metric. We implemented ReplayAR on a Hololens 2 MR headset and used it to evaluate open and closed-source pose estimation methods on standard datasets and our own collected traces. The results suggest that Visual Difference better reflects what is shown on the MR display compared to ATE. We hope that ReplayAR can encourage reproducible evaluation of world-locked MR, and towards this, we release the open-source code.

Recent successes in natural language processing have led to the proliferation of large language models (LLMs) by multiple providers. Each LLM offering has different inference accuracy, monetary cost, and latency, and their accuracy further depends on the exact wording of the question (i.e., the specific prompt). At the same time, users often have a limit on monetary budget and latency to answer all their questions, and they do not know which LLMs to choose for each question to meet their accuracy and long term budget requirements. To navigate this rich design space, we propose TREACLE (Thrifty Reasoning via Context-Aware LLM and Prompt Selection), a reinforcement learning policy that jointly selects the model and prompting scheme while respecting the user's monetary cost and latency constraints. TREACLE uses the problem context, including question text embeddings (reflecting the type or difficulty of a query) and the response history (reflecting the consistency of previous responses) to make smart decisions. Our evaluations on standard reasoning datasets (GSM8K, CSQA, and LLC) with various LLMs and prompts show that TREACLE enables cost savings of up to 85% compared to baselines, while maintaining high accuracy. Importantly, it provides the user with the ability to gracefully trade off accuracy for cost.

The attention mechanism within the transformer architecture enables the model to weigh and combine tokens based on their relevance to the query. While self-attention has enjoyed major success, it notably treats all queries $q$ in the same way by applying the mapping $V^\top\text{softmax}(Kq)$, where $V,K$ are the value and key embeddings respectively. In this work, we argue that this uniform treatment hinders the ability to control contextual sparsity and relevance. As a solution, we introduce the Selective Self-Attention (SSA) layer that augments the softmax nonlinearity with a principled temperature scaling strategy. By controlling temperature, SSA adapts the contextual sparsity of the attention map to the query embedding and its position in the context window. Through theory and experiments, we demonstrate that this alleviates attention dilution, aids the optimization process, and enhances the model's ability to control softmax spikiness of individual queries. We also incorporate temperature scaling for value embeddings and show that it boosts the model's ability to suppress irrelevant/noisy tokens. Notably, SSA is a lightweight method which introduces less than 0.5\% new parameters through a weight-sharing strategy and can be fine-tuned on existing LLMs. Extensive empirical evaluations demonstrate that SSA-equipped models achieve a noticeable and consistent accuracy improvement on language modeling benchmarks.

WiFi-enabled Power Line Communications (PLC) range extenders can extend coverage in homes and enterprises. However, a dense deployment of a large number of PLC extenders in enterprise settings can cause an inefficient sharing of the PLC capacity, where many extenders contend for a share of access to the backhaul network comprising of the electrical wiring (power lines), thereby drastically impacting any gains from using these extenders on the wireless part of the network. In this paper, we address this issue by developing a framework, Priza, for clustering the WiFi-PLC extenders to intelligently form a DAS (distributed antenna system) to mitigate the inefficiency of sharing the PLC backhaul. By appropriately managing clustering and reuse, Priza improves the PLC backhaul sharing, while at the same time, harnessing the power pooling and diversity gains from DAS on the wireless part of the network, to boost user throughputs. We evaluate Priza via real testbed experiments and high-fidelity simulations and demonstrate that it can increase the aggregate throughput by up to 131.5% over the non-DAS reuse baseline, 74% over the best DAS baseline that constructs equally-sized DAS cells based on extender proximity, and 331.3% over a greedy DAS baseline that creates as large DAS cells as possible.

World-scale augmented reality (AR) is a form of AR where users move around the real world, viewing and interacting with 3D models at specific locations. However, given the geographical scale of world-scale AR, pre-fetching and storing numerous high-quality 3D models locally on the device is infeasible. For example, it would be impossible to download and store 3D ads from all the storefronts in a city onto a single device. A key challenge is thus deciding which remotely-stored 3D models should be fetched onto the AR device from an edge server, in order to render them in a timely fashion - yet with high visual quality - on the display. In this work, we propose a 3D model retrieval framework that makes intelligent decisions of which quality of 3D models to fetch, and when. The optimization decision is based on quality-compression tradeoffs, network bandwidth, and predictions of which 3D models the AR user is likely to view next. To support our framework, we collect real-world traces of AR users playing a world-scale AR game, and use this to drive our simulation and prediction modules. Our results show that the proposed framework can achieve higher visual quality of the 3D models while missing fewer display deadlines (by 20%) and wasting fewer bytes (by 10x), compared to a baseline approach of pre-fetching models within a fixed distance of the user.

Augmented reality (AR) devices perform visual simultaneous localization and mapping (SLAM) to map the real world and localize themselves in it, enabling them to render the virtual holograms appropriately. Current multi-user AR platforms fall short in that they only allow asymmetric sharing of this SLAM information, resulting in multiple "secondary" devices viewing holograms placed by a single "primary" device, instead of equal participation. The goal of this work is to enable all AR devices to participate equally, by constructing a common global map to which all AR devices can contribute. However, doing so with low latency and high accuracy is challenging on resource-constrained mobile devices. This work proposes an appropriate partitioning between clients and a server to achieve high-throughput, low latency, multi-user SLAM. In our system, SLAM-Share, the edge server performs the complex SLAM computations so that the client devices need only perform lightweight operations. The server utilizes shared memory and efficient map merging to build and update a global map from different clients. It also exploits the parallelism of GPU processing to achieve high-performance tracking. Evaluations show that SLAM-Share is able to achieve significant tracking speedups (up to 50% reduction compared to alternative approaches), maintain good localization accuracy, and merge and update maps within 200 ms.

Contact tracing is a key mechanism to help contain the COVID-19 pandemic and other pandemics in the future. In this work, we propose using 5G channel signatures – specifically, mm-Wave channel signatures – to perform contact tracing and infer early sources of the infection. Our network-side approach is motivated by the density of mm-Wave base stations, coupled with the large amount of data about mobile device signals already being collected by cellular operators. We model the contact tracing problem as a graph mining problem, and develop machine learning models to estimate contacts between UEs based on 5G channel signatures such as received power. These contacts are also used to infer the original sources of the infection. Simulations of our proposed method using the ns-3 5G mmWave module suggest that contact can be inferred with a recall of 85% and specificity of 94%. Our infection sources estimation method can accurately rank the most likely infection sources, with the true infection sources lying in the top 25% of the ranked list on average. These methods represent a first step towards network-based contact tracing, and can complement other contact tracing methods to help reduce the spread of disease.

Collaborative AR applications are gaining popularity, but have heavy computing requirements for identifying and tracking AR devices and objects in the ecosystem. Prior AR frameworks typically rely on edge infrastructure to offload AR's compute-heavy tasks. However, such infrastructure may not always be available, and continuously running AR computations on user devices can rapidly drain battery and impact application longevity. In this work, we enable infrastructure-free mobile AR with a low energy footprint, by using collaborative time slicing to distribute compute-heavy AR tasks across user devices. Realizing this idea is challenging because distributed execution can result in inconsistent synchronization of the AR virtual overlays. Our framework, FreeAR, tackles this with novel lightweight techniques for tightly synchronized virtual overlay placements across user views, and low latency recovery upon disruptions. We prototype FreeAR on Android and show that it can improve the virtual overlay positioning accuracy (with respect to the IOU metric) by up to 78%, relative to state-of-the-art collaborative AR systems, while also reducing power by up to 60% relative to a direct application of those prior solutions.

In augmented reality (AR), virtual objects can drift away from their original intended locations, significantly impairing a user’s experience. Traditionally, a virtual object’s drift is approximated by the device localization drift, which is measured using specialized hardware such as 3D scanners or laser-based positioning systems. However, with AR rapidly becoming more popular, there is a need for a lightweight, software-based approach to evaluate the drift of virtual objects. This software should be easy for researchers and developers to use, without requiring specialized hardware or extensive environment setup.Towards this, this paper presents RealityCheck, an opensource AR evaluation tool that reports the drift of AR virtual objects in the world coordinate system, requiring only paper printouts and minimal modifications to the AR app. RealityCheck is designed to measure the drift of a virtual object across time of a single user, as well as the positioning differences of the same virtual objects as seen by multiple users. Our prototype is implemented on an Android smartphone running the ARCore platform, and evaluated in indoor and outdoor scenarios under a variety of user mobility patterns with traces of different lengths. We compared the results of RealityCheck with the ground truth position of the virtual object, and showed that RealityCheck matches the ground truth within 1.5 cm on average.

WiFi-capable PLC (Poiver Line Communications) plug-and-play extenders are becoming popular to improve WiFi range and coverage in homes and enterprises. As shown in prior work, unlike an Ethernet backhaul, the PLC backhaul may not support high data rates. In addition, clients (users) that are either far or partially occluded from the WiFi-PLC extender they associate with can experience fading and shadowing, which degrades the throughput on the wireless link. Thus, both the PLC and WiFi backhauls will influence a user’s end-to-end throughput. In this paper, we seek to exploit the presence of multiple PLC extenders that may be plugged in, by combining their transmissions in a distributed antenna system (DAS), to boost client throughputs in a home setting. Specifically, we design PLC-DAS to determine which PLC extenders are the best candidates for forming a joint DAS transmitter cluster to each client. PLC-DAS is designed based on a real measurement study and not only accounts for the WiFi link qualities from the extenders to the users, but also the PLC link qualities from each extender to a master router which is typically deployed in homes. PLC-DAS is flexible and can maximize the throughput under different fairness objectives. We evaluate PLC-DAS via extensive simulations and show that it can increase the aggregate throughput by up to 4.5x compared to blindly using all WiFi PLC extenders to form a DAS transmitter, while maintaining a fairness Jain’s index value of at least 0.97 with proportional and max-min fairness models.

Imbalanced datasets are commonplace in modern machine learning problems. The presence of under-represented classes or groups with sensitive attributes results in concerns about generalization and fairness. Such concerns are further exacerbated by the fact that large capacity deep nets can perfectly fit the training data and appear to achieve perfect accuracy and fairness during training, but perform poorly during test. To address these challenges, we propose AutoBalance, a bi-level optimization framework that automatically designs a training loss function to optimize a blend of accuracy and fairness-seeking objectives. Specifically, a lower-level problem trains the model weights, and an upper-level problem tunes the loss function by monitoring and optimizing the desired objective over the validation data. Our loss design enables personalized treatment for classes/groups by employing a parametric cross-entropy loss and individualized data augmentation schemes. We evaluate the benefits and performance of our approach for the application scenarios of imbalanced and group-sensitive classification. Extensive empirical evaluations demonstrate the benefits of AutoBalance over state-of-the-art approaches. Our experimental findings are complemented with theoretical insights on loss function design and the benefits of the train-validation split. All code is available open-source.

Multi-user augmented reality (AR), where multiple co-located users view a common set of virtual objects, is becoming increasingly popular. For example, Google Just a Line allows multiple users to draw virtual graffiti in the same physical space. Multi-user AR requires network communications in order to coordinate the positions of the virtual objects on each user's display, yet there is currently little understanding of how such apps communicate. In this work, we address this key gap in knowledge by showing that the communicated data directly impacts the latency and positioning of the virtual objects rendered on the users' displays. We develop solutions to these problems that we find along three facets: (1) efficient communication strategies that trade off communication latency for spatial consistency of the virtual objects; (2) a new metric that enables mobile AR devices to update their virtual objects as they move around and observe more of the scene; and (3) a tool to automatically quantify how much the virtual objects' positions inadvertently change in time and space. Our evaluation is performed on Android smartphones running open-source AR. The results show that our system, SPAR, can decrease the latency by up to 55%, while decreasing the spatial inconsistency by up to 60%, compared to baseline methods.

Predicting a user's field-of-view (FoV) accurately can help to significantly reduce the high bandwidth requirements for 360° video streaming, as it enables sending only the tiles corresponding to the predicted FoV. Since many approaches for user head-orientation (i.e., FoV) prediction have been proposed in the literature, ranging from simple linear regression to more complex neural networks, it is difficult to comprehensively decide which method to use. Towards resolving this gap in knowledge, in this work we benchmark user prediction algorithms over an aggregation of multiple datasets and study the implications of this analysis. Our results demonstrate that it is indeed difficult for any prediction algorithm to accurately predict a user's FoV beyond a very short future time window of approximately 300 ms. We also observe that users' viewing behavior is dominated by sideways head movement, rather than up-and-down. These findings have implications on network bandwidth, latency, and playback buffering at the client: (1) Extra "padding" tiles are needed around the user's FoV in order to correct for prediction errors; in particular, a rectangular padding achieves lower stall rate than square padding, for the same bandwidth usage; (2) Video playout buffers, network delay, and jitter need to be small in order to avoid stale predictions of the user's field-of-view, which are only valid 300 ms into the future; (3) Per-video and per-user personalization of the padding can save bandwidth for slow-moving users or videos. We mathematically quantify these tradeoffs and present simulation results to demonstrate these findings and implications. Our results have implications for FoV prediction methods in future 360°streaming systems.

Power Line Communication (PLC) based WiFi extenders can improve WiFi coverage in homes and enterprises. Unlike in traditional WiFi networks which use an underlying high data rate Ethernet backhaul, a PLC backhaul may not support high data rates. Specifically, our measurements show that arbitrarily affiliating users to PLC-WiFi extenders or based on their WiFi channel qualities alone may lead to poor network performance due to the differences in PLC link capacities. Thus, in this paper we build a framework, WOLT, to solve the problem of assigning users to the appropriate PLC-WiFi extenders to increase the aggregate network throughput in an enterprise setting, where one may expect a relatively large number of power outlets. WOLT accounts for both the qualities of the two concatenated links viz., the PLC and WiFi links. It hinges on estimating the best capacity offered by the PLC links, and accounting for these while assigning users. It incorporates a polynomial-time algorithm that assigns only a subset of the users to maximize the aggregate throughput on the PLC links, and then assigns the remaining users such that the degradation in the aggregate throughput is minimized. WOLT is evaluated through simulations and real testbed experiments with commodity PLCWiFi extenders, and improves aggregate throughput by more than 2.5× compared to a greedy user association baseline.

Augmented reality (AR) apps where multiple users interact within the same physical space are gaining in popularity (e.g., shared AR mode in Pokemon Go, virtual graffiti in Google's Just a Line). However, multi-user AR apps running over the cellular network can experience very high end-to-end latencies (measured at 12.5 s median on a public LTE network). To characterize and understand the root causes of this problem, we perform a first-of-its-kind measurement study on both public LTE and industry LTE testbed for two popular multi-user AR applications, yielding several insights: (1) The radio access network (RAN) accounts for a significant fraction of the end-to-end latency (31.2%, or 3.9 s median), resulting in AR users experiencing high, variable delays when interacting with a common set of virtual objects in off-the-shelf AR apps; (2) AR network traffic is characterized by large intermittent spikes on a single uplink TCP connection, resulting in frequent TCP slow starts that can increase user-perceived latency; (3) Applying a common traffic management mechanism of cellular operators, QoS Class Identifiers (QCI), can help by reducing AR latency by 33% but impacts non-AR users. Based on these insights, we propose network-aware and network-agnostic AR design optimization solutions to intelligently adapt IP packet sizes and periodically provide information on uplink data availability, respectively. Our solutions help ramp up network performance, improving the end-to-end AR latency and goodput by ~40-70%.

Growing mobile data usage has led to end users paying substantial data costs, while Internet service providers (ISPs) struggle to upgrade their networks to keep up with demand and maintain high quality-of-service (QoS). This problem is particularly severe for smaller ISPs with less capital. Instead of simply upgrading their network infrastructure, ISPs can pool their networks to provide a good QoS and attract more users. Such a vISP (virtual ISP), for example, Google's Project Fi, allows users to access any of its partner ISPs' networks. We provide the first systematic analysis of a vISP's economic impact, showing that the vISP provides a viable solution for smaller ISPs attempting to attract more users, but may not maintain a positive profit if users' data demands evolve. To do so, we consider users' decisions of whether to defect from their current ISP to the vISP, as well as existing ISPs' decisions on whether to partner with the vISP. We derive the vISP's dependence on user behavior and partner ISPs: users with very light or very heavy usage are the most likely to defect, while ISPs with heavy-usage customers can benefit from declining to partner with the vISP. Our analytical results are verified with extensive numerical simulations.

Augmented reality is an emerging application on mobile devices. However, there is a lack of understanding of the communication requirements and challenges of multi-user AR scenarios. In this position paper, we propose several important research issues that need to be addressed for low-latency, accurate shared AR experiences: (a) Systems tradeoffs of AR communication architectures used today in mobile AR platforms; (b) Understanding AR communication patterns and adapting the AR application layer to dynamically changing network conditions; and (c) Tools and methodologies to evaluate AR quality of experience in real time on mobile devices. We present preliminary measurements of off-the-shelf mobile AR platforms as well as results from our AR system, ShareAR, illustrating performance tradeoffs and indicating promising new research directions.

For various applications, the relations between the dependent and independent variables are highly nonlinear. Consequently, for large scale complex problems, neural networks and regression trees are commonly preferred over linear models such as Lasso. This work proposes learning the feature nonlinearities by binning feature values and finding the best fit in each quantile using non-convex regularized linear regression. The algorithm first captures the dependence between neighboring quantiles by enforcing smoothness via piecewise-constant/linear approximation and then selects a sparse subset of good features. We prove that the proposed algorithm is statistically and computationally efficient. In particular, it achieves linear rate of convergence while requiring near-minimal number of samples. Evaluations on real datasets demonstrate that algorithm is competitive with current state-of-the-art and accurately learns feature nonlinearities.

Accurate tracking of objects in the real world is highly desirable in Augmented Reality (AR) to aid proper placement of virtual objects in a user's view. Deep neural networks (DNNs) yield high precision in detecting and tracking objects, but they are energy-heavy and can thus be prohibitive for deployment on mobile devices. Towards reducing energy drain while maintaining good object tracking precision, we develop a novel software framework called MARLIN. MARLIN only uses a DNN as needed, to detect new objects or recapture objects that significantly change in appearance. It employs lightweight methods in between DNN executions to track the detected objects with high fidelity. We experiment with several baseline DNN models optimized for mobile devices, and via both offline and live object tracking experiments on two different Android phones (one utilizing a mobile GPU), we show that MARLIN compares favorably in terms of accuracy while saving energy significantly. Specifically, we show that MARLIN reduces the energy consumption by up to 73.3% (compared to an approach that executes the best baseline DNN continuously), and improves accuracy by up to 19× (compared to an approach that infrequently executes the same best baseline DNN). Moreover, while in 75% or more cases, MARLIN incurs at most a 7.36% reduction in location accuracy (using the common IOU metric), in more than 46% of the cases, MARLIN even improves the IOU compared to the continuous, best DNN approach.

Deep learning is currently widely used in a variety of applications, including computer vision and natural language processing. End devices, such as smartphones and Internet-of-Things sensors, are generating data that need to be analyzed in real time using deep learning or used to train deep learning models. However, deep learning inference and training require substantial computation resources to run quickly. Edge computing, where a fine mesh of compute nodes are placed close to end devices, is a viable way to meet the high computation and low-latency requirements of deep learning on edge devices and also provides additional benefits in terms of privacy, bandwidth efficiency, and scalability. This paper aims to provide a comprehensive review of the current state of the art at the intersection of deep learning and edge computing. Specifically, it will provide an overview of applications where deep learning is used at the network edge, discuss various approaches for quickly executing deep learning inference across a combination of end devices, edge servers, and the cloud, and describe the methods for training deep learning models across multiple edge devices. It will also discuss open challenges in terms of systems performance, network technologies and management, benchmarks, and privacy. The reader will take away the following concepts from this paper: understanding scenarios where deep learning at the network edge can be useful, understanding common techniques for speeding up deep learning inference and performing distributed training on edge devices, and understanding recent trends and opportunities.

Deep learning shows great promise in providing more intelligence to augmented reality (AR) devices, but few AR apps use deep learning due to lack of infrastructure support. Deep learning algorithms are computationally intensive, and front-end devices cannot deliver sufficient compute power for real-time processing. In this work, we design a framework that ties together front-end devices with more powerful backend “helpers” (e.g., home servers) to allow deep learning to be executed locally or remotely in the cloud/edge. We consider the complex interaction between model accuracy, video quality, battery constraints, network data usage, and network conditions to determine an optimal offloading strategy. Our contributions are: (1) extensive measurements to understand the tradeoffs between video quality, network conditions, battery consumption, processing delay, and model accuracy; (2) a measurement-driven mathematical framework that efficiently solves the resulting combinatorial optimization problem; (3) an Android application that performs real-time object detection for AR applications, with experimental results that demonstrate the superiority of our approach.

Growing mobile data usage has led to end users paying substantial data costs, while Internet service providers (ISPs) struggle to upgrade their networks to keep up with demand and maintain high quality-of-service (QoS). This problem is particularly severe for smaller ISPs with less capital. Instead of simply upgrading their network infrastructure, ISPs can pool their networks to provide a good QoS and attract more users. Such a vISP (virtual ISP), for example, Google's Project Fi, allows users to access any of its partner ISPs' networks. We provide the first systematic analysis of a vISP's economic impact, showing that the vISP provides a viable solution for smaller ISPs attempting to attract more users, but may not maintain a positive profit if users' data demands evolve. To do so, we consider users' decisions of whether to defect from their current ISP to the vISP, as well as ISPs' decisions on whether to partner with the vISP. We derive the vISP's dependence on user behavior and partner ISPs: users with very light or very heavy usage are the most likely to defect, while ISPs with heavy-usage customers can benefit from declining to partner with the vISP. Our analytical results are verified with extensive numerical simulations.

With growing data demand and the current dearth of spectrum, mobile operators are looking for new frequency bands to satisfy data-hungry users. One promising avenue of expansion is TV white spaces, which are currently available to secondary users as long as they do not interfere with primary (i.e., incumbent) users. In this work, we explore the benefits of offloading cellular traffic onto TV white spaces. We develop an analytical model and efficient algorithms to assign users to the cellular network or white space channels by considering their channel gains, multi-user interference on white space channels, and the cost of switching between different networks. We perform extensive data-driven simulations in two representative urban scenarios based on publicly available datasets. Our results show that white spaces can increase capacity by 16-62%, depending on the environment, but careful network selection is necessary to ensure that maximum capacity gains are realized. Moreover, we show that white spaces provide a significant benefit in serving indoor users where cellular channel conditions are poor. Specifically, our algorithms can offload up to 40% of cellular traffic to white spaces for indoor scenarios.

Internet service providers (ISPs) struggle to invest in upgrading their networks to catch up with growing mobile data demand, while users have to face significant data overage fees. Pooling ISPs' network infrastructures can potentially enable better user experience and lower prices. For example, Google recently launched a cross-carrier MVNO (mobile virtual network operator) data plan called Project Fi, where users' devices can automatically access either of two partner cellular networks or any available open WiFi network. We consider the economic impact of cross-carrier MVNOs on the mobile data market. We begin by analyzing a network selection strategy that optimizes cross-carrier users' costs. We then study ISPs' behavior, deriving the prices that partner ISPs charge the cross-carrier MVNO and that the cross-carrier MVNO charges its end users. Although the cross-carrier MVNO may lose money from selling data, it can offset this loss with side revenue, e.g., advertisement revenue when users consume more content. We derive conditions under which the cross-carrier MVNO achieves a profit and its users reduce their costs. Finally, we use a real-world network quality dataset to simulate users' network selection behavior and demonstrate the benefits of the ISP competition brought by the cross-carrier MVNO.

Deploying caches in mobile networks, especially in the radio access network (RAN) is regarded as a promising way to improve mobile user experiences and alleviate the increasing pressure of traffic growth. However, the characteristics of mobile traffic and the performance of RAN caching still remains unclear. In this paper, we extensively analyze the traffic characteristics, the content popularity and the cache performance using a unique dataset collected from a commercial LTE network of China Mobile, from the perspective of mobile access network. The dataset spans nearly a week and consists of a collection of approximately 62.1 millions HTTP sessions, generated by more than 3200 users distributed across three base stations. Based on this realistic dataset, we observe that HTTP traffic can be reduced by 24.4% on average and the hit ratio can reach up to 42.2%, using 100GB cache size. The implications on some fundamental design issues of practical RAN caching systems, including reasonable size of RAN cache, suitable locations of cache deployment and potential benefits of collaborative RAN caching, are further presented. We believe our findings will shed light on practical RAN caching system design.

With recent standardization and deployment of LTE eMBMS, cellular multicast is gaining traction as a method of efficiently using wireless spectrum to deliver large amounts of multimedia data to multiple cell sites. Cellular operators still seek methods of performing optimal resource allocation in eMBMS based on a complete understanding of the complex interactions among a number of mechanisms: the multicast coding scheme, the resources allocated to unicast users and their scheduling at the base stations, the resources allocated to a multicast group to satisfy the user experience of its members, and the number of groups and their membership, all of which we consider in this work. We determine the optimal allocation of wireless resources for users to maximize proportional fair utility. To handle the heterogeneity of user channel conditions, we efficiently and optimally partition multicast users into groups so that users with good signal strength do not suffer by being grouped together with users of poor signal strength. Numerical simulations are performed to compare our scheme to practical heuristics and state-of-the-art schemes. We demonstrate the tradeoff between improving unicast user rates and improving spectrum efficiency through multicast. Finally, we analyze the interaction between the globally fair solution and individual user's desire to maximize its rate. We show that even if the user deviates from the global solution in a number of scenarios, we can bound the number of selfish users that will choose to deviate.

The recently proposed 3-Tier access model for Whitespace by the Federal Communications Commission (FCC) mandates certain classes of devices to share frequency bands in space and time. These devices are envisioned to be a heterogeneous mixture of licensed (Tier-1 and Tier-2) and unlicensed, opportunistic devices (Tier-3). The hierarchy in accessing the channel calls for superior adaptation of Tier-3 devices with varying spectral opportunity. While policies are being ratified for efficient sharing, it also calls for redesigning many common applications to adapt to this novel paradigm. In this paper, we focus on the ever-increasing demand for video streaming and present a methodology suitable for Tier-3 devices in the shared access model. Our analysis begins with a stress test of commonly adopted video streaming methods under the new sharing model. This is followed by the design of a robust MDP-based solution that proactively adapts to fast-varying channel conditions, providing better user quality of experience when compared to existing solutions, such as MPEG-DASH. We evaluate our solution on an experimental testbed and find that our MDP-based algorithm outperforms DASH, and partial information of Tier-2 dynamics improves video quality.

As the growth of mobile video traffic outpaces that of cellular network speed, industry is adopting HTTP-based adaptive video streaming technology which enables dynamic adaptation of video bit-rates to match changing network conditions. However, recent measurement studies have observed problems in fairness, stability, and efficiency of resource utilization when multiple adaptive video flows compete for bandwidth on a shared wired link. Through experiments and simulations, we confirm that such undesirable behavior manifests itself in cellular networks as well. To overcome these problems, we design an in-network resource management framework, AVIS, that schedules HTTP-based adaptive video flows on cellular networks. AVIS effectively manages the resources of a cellular base station across adaptive video flows. AVIS also provides a framework for mobile operators to achieve a desired balance between optimal resource allocation and user quality of experience. AVIS has three key differentiating features: (1) It optimally computes the bit-rate allocation for each user, (2) It includes a scheduler and per-flow shapers to enforce bit-rate stability of each flow and (3) It leverages the resource virtualization technique to separate resource management of adaptive video flows from regular video flows. We implement a prototype system of AVIS and evaluate it on both a WiMAX network testbed and a LTE system simulator to show its efficacy and scalability.

Two emerging trends of Internet applications, video traffic becoming dominant and usage-based pricing becoming prevalent, are at odds with each other. Given this conflict, is there a way for users to stay within their monthly data plans (data quotas) without suffering a noticeable degradation in video quality? In this work, we develop an online video adaptation system, called Quota Aware Video Adaptation (QAVA), that manages this tradeoff by leveraging the compressibility of videos and by predicting consumer usage behavior throughout a billing cycle. We propose the QAVA architecture and develop its main modules, including Stream Selection, User Profiling, and Video Profiling. Online algorithms are designed through dynamic programming and evaluated using real video request traces. Empirical results suggest that QAVA can provide an effective solution to the dilemma of usage-based pricing of heavy video traffic.

With the convergence of ultra-low-power communications and energy-harvesting technologies, networking self-sustainable ubiquitous devices is becoming feasible. Hence, we have been recently developing new devices, referred to as Energy Harvesting Active Networked Tags (EnHANTs). These small, flexible, and energetically self-reliant tags can be seen as a new class of devices in the domain between RFIDs and sensor networks. EnHANTs are made possible by advances in ultra-lowpower ultra-wideband (UWB) communications and in organic semiconductor-based energy harvesting materials. They will enable novel tracking applications, such as continuous monitoring of objects and locating misplaced items. In this demo, we present phase I EnHANT prototypes. These prototypes are much larger than the envisioned EnHANTs and do not include custom-made UWB and organic electronic components. Yet, they serve as platforms for preliminary experiments and allow demonstrating energy harvesting-adaptive EnHANT communications. Each prototype is based on a MICA2 mote and includes a custom-designed sensor board with a light sensor and a solar cell, which are used to determine the light energy received from the environment. We have also designed a monitoring system which is used in the demo to show how the EnHANT prototypes adjust their communications patterns based on their energy harvesting parameters.