MobiCom 2019 Papers

People use their mobile devices anywhere and anytime to run various apps, and the information shown on their device screens can be seen by nearby (unauthorized) parties, called shoulder surfers. To mitigate this privacy threat, we have developed HideScreen by utilizing the human vision and optical system properties to hide the users' on-screen information from the shoulder surfers.

IoT and deep learning based computer vision together create an immense market opportunity, but running deep neural networks (DNNs) on resource-constrained IoT devices remains challenging. Offloading DNN inference to an edge server is a promising solution, but limited wireless bandwidth bottlenecks its end-to-end performance and scalability. While IoT devices can adopt source compression to cope with the limited bandwidth, existing compression algorithms (or codecs) are not designed for DNN (but for human eyes), and thus, suffer from either low compression rates or high DNN inference errors. This paper presents GRACE, a DNN-aware compression algorithm that facilitates the edge inference by significantly saving the network bandwidth consumption without disturbing the inference performance. Given a target DNN, GRACE (i) analyzes DNN's perception model w.r.t both spatial frequencies and colors and (ii) generates an optimized compression strategy for the model -- one-time offline process. Next, GRACE deploys thus-generated compression strategy at IoT devices (or source) to perform online source compression within the existing codec framework, adding no extra overhead. We prototype GRACE on JPEG (the most popular image codec framework), and our evaluation results show that GRACE indeed achieves the superior compression performance over existing strategies for key DNN applications. For semantic segmentation tasks, GRACE reduces a source size by 23% compared to JPEG with similar interference accuracy (0.38% lower than GRACE). Further, GRACE even achieves 7.5% higher inference accuracy than JPEG with a commonly used quality level of 75 does. For classification tasks, GRACE reduces the bandwidth consumption by 90% over JPEG with the same inference accuracy.

MP-H2 is a client-only, HTTP-based multipath solution. It enables an HTTP client to fetch content (an HTTP object) over multiple network paths such as WiFi and cellular on smartphones. Compared to MPTCP, MP-H2 offers several key advantages including server transparency, middlebox compatibility, and friendliness to CDN, anycast, and load balancing. MP-H2 strategically splits the file into byte range requests sent over multipath, and dynamically balances the workload across all paths. Furthermore, MP-H2 leverages new features in HTTP/2 including stream multiplexing, flow control, and application-layer PING to boost the performance. MP-H2 also supports multi-homing where each path contacts a different CDN server for enhanced performance. Evaluations show that MP-H2 offers only slightly degraded performance (6% on average) while being much easier to deploy compared to MPTCP. Compared to other state-of-the-art HTTP multipath solutions, MP-H2 reduces the file download time by up to 47%, and increases the DASH video streaming bitrate by up to 44%.

Device Administrator (DA) capabilities for mobile devices, e.g., remote locking/wiping, or enforcing password strength, were originally introduced to help organizations manage phone fleets or enable parental control. However, DA capabilities have been subverted and abused: malicious apps have used DA to create ransomware or lock users out, while benign apps have used DA to prevent or hinder uninstallation; in certain cases the only remedy is to factory-reset the phone. We call these apps "Deathless Device Administrator" (DDA), i.e., apps that cannot be uninstalled. We provide the first systematic study of Android DA capabilities, DDA apps, DDA-attack resistance across Android versions, and DDA-induced families in malicious apps. To enable scalable studies of questionable DA behavior, we developed DAAX, a static analyzer which exposes potential DA abuse effectively and efficiently. In a corpus of 39,459 apps (20,467 malicious and 18,992 benign) DAAX has found 4,135 DA apps and 691 potential DDA apps. The static analysis results on the 4,135 apps were cross-checked via dynamic analysis on at least 3 phones, confirming 578 true DDAs, including apps currently on Google Play. The study has shown that DAAX is effective (84.8% F-measure) and efficient (analysis typically takes 205 seconds per app).

Despite the pervasive use of real-time video telephony services, the users' quality of experience (QoE) remains unsatisfactory, especially over the mobile Internet. Previous work studied the problem via controlled experiments, while a systematic and in-depth investigation in the wild is still missing. To bridge the gap, we conduct a large-scale measurement campaign on \appname, an operational mobile video telephony service. Our measurement logs fine-grained performance metrics over 1 million video call sessions. Our analysis shows that the application-layer video codec and transport-layer protocols remain highly uncoordinated, which represents one major reason for the low QoE. We thus propose \name, a machine learning based framework to resolve the issue. Instead of blindly following the transport layer's estimation of network capacity, \name reviews historical logs of both layers, and extracts high-level features of codec/network dynamics, based on which it determines the highest bitrates for forthcoming video frames without incurring congestion. To attain the ability, we train \name with the aforementioned massive data traces using a custom-designed imitation learning algorithm, which enables \name to learn from past experience. We have implemented and incorporated \name into \appname. Our experiments show that \name outperforms state-of-the-art solutions, improving video quality while reducing stalling time by multi-folds under various practical scenarios.

The automotive industry is increasingly employing software- based solutions to provide value-added features on vehicles, especially with the coming era of electric vehicles and autonomous driving. The ever-increasing cyber components of vehicles (i.e., computation, communication, and control), however, incur new risks of anomalies, as demonstrated by the millions of vehicles recalled by different manufactures. To mitigate these risks, we design B-Diag, a battery-based diagnostics system that guards vehicles against anomalies with a cyber-physical approach, and implement B-Diag as an add-on module of commodity vehicles attached to automotive batteries, thus providing vehicles an additional layer of protection. B-Diag is inspired by the fact that the automotive battery operates in strong dependency with many physical components of the vehicle, which is observable as correlations between battery voltage and the vehicle's corresponding operational parameters, e.g., a faster revolutions-per-minute (RPM) of the engine, in general, leads to a higher battery voltage. B-Diag exploits such physically-induced correlations to diagnose vehicles by cross-validating the vehicle information with battery voltage, based on a set of data-driven norm models constructed online. Such a design of B-Diag is steered by a dataset collected with a prototype system when driving a 2018 Subaru Crosstrek in real-life over 3 months, covering a total mileage of about 1, 400 miles. Besides the Crosstrek, we have also evaluated B-Diag with driving traces of a 2008 Honda Fit, a 2018 Volvo XC60, and a 2017 Volkswagen Passat, showing B-Diag detects vehicle anomalies with >86% (up to 99%) averaged detection rate.

Wi-Fi localization and tracking face accuracy limitations dictated by antenna count (for angle-of-arrival methods) and frequency bandwidth (for time-of-arrival methods). This paper presents mD-Track, a device-free Wi-Fi tracking system capable of jointly fusing information from as many dimensions as possible to overcome the resolution limit of each individual dimension. Through a novel path separation algorithm, mD-Track can resolve multipath at a much finer-grained resolution, isolating signals reflected off targets of interest. mD-Track can localize human passively at a high accuracy with just a single Wi-Fi transceiver pair. mD-Track also introduces novel methods to greatly streamline its estimation algorithms, achieving real-time operation. We implement mD-Track on both WARP and cheap off-the-shelf commodity Wi-Fi hardware, and evaluate its performance in different indoor environments.

Understanding spectrum characteristics with little prior knowledge requires fine-grained spectrum data in the frequency, spatial, and temporal domains; gathering such a diverse set of measurements results in a large data volume. Analysis of the resulting dataset poses unique challenges; methods in the status quo are tailored for specific spectrum-related applications (apps), and are ill equipped to process data of this magnitude. In this paper, we design BigSpec, a general-purpose framework that allows for fast processing of apps. The key idea is to reduce computation costs by performing computation extensively on compressed data that preserves signal features. Adhering to this guideline, we build solutions for three apps, i.e., energy detection, spatio-temporal spectrum estimation, and anomaly detection. These apps were chosen to highlight BigSpec's efficiency, scalability, and extensibility. To evaluate BigSpec's performance, we collect more than 1 terabyte of spectrum data spanning a year, across 300MHz-4GHz, covering 400 km2. Compared with baselines and prior works, we achieve 17× run time efficiency, sublinear rather than linear run time scalability, and extend the definition of anomaly to different domains (frequency & spatio-temporal). We also obtain high-level insights from the data to provide valuable advice on future spectrum measurement and data analysis.

Longer range, in rural/urban IoT networks, allow a large geographical coverage with only a few base-stations, making their deployment and operation economical. In this paper we explore the 150-174 MHz spectrum for long range IoT networks comprising unlicensed MURS and licensed VHF narrowbands. Range in these bands is boosted by the lower RF frequencies as well as higher transmit powers allowed by the FCC. Through a 400 sq km wide area deployment study, we show that, these spectrum bands can provide > 20× the geographical coverage than that in the 900 MHz ISM band LoRa. Increased range translates to greater uplink IoT device traffic. The key contribution of this paper is a novel technique - Blind Distributed MU-MIMO, that allows capacity to scale with the number of antennas (base- stations) while not requiring any coordinated channel measurements between the devices and IoT base- stations. This requirement is crucial since in IoT networks power constrained IoT devices typically sleep and wake up to transmit short messages in response to unpredictable events without any coordination with the base-stations. We demonstrate the efficacy of Blind Distributed MU-MIMO through a real wide area deployment.

Parkinson's disease (PD) is a chronic neurodegenerative disorder resulting from the progressive loss of dopaminergic nerve cells. People with PD usually demonstrate deficits in performing basic daily activities, and the relevant annual social cost can reach about $25 billion in the United States. Early detection of PD plays an important role in symptom relief and improvement in the performance of activities in daily life (ADL), which eventually reduces societal and economic burden. However, conventional PD detection methods are inconvenient in daily life (e.g., requiring users to wear sensors). To overcome this challenge, we propose and identify the non-speech body sounds as the new PD biomarker, and utilize the data in smartphone usage to realize the passive PD detection in daily life without interrupting the user. Specifically, we present PDVocal, an end-to-end smartphone-based privacy-preserving system towards early PD detection. PDVocal can passively recognize the PD digital biomarkers in the voice data during daily phone conversation. At the user end, PDVocal filters the audio stream and only extracts the non-speech body sounds (e.g., breathing, clearing throat and swallowing) which contain no privacy-sensitive content. At the cloud end, PDVocal analyzes the body sounds of interest and assesses the health condition using a customized residual network. For the sake of reliability in real-world PD detection, we investigate the method of the performance optimizer including an opportunistic learning knob and a long-term tracking protocol. We evaluate our proposed PDVocal on a collected dataset from 890 participants and real-life conversations from publicly available data sources. Results indicate that non-speech body sounds are a promising digital biomarker for privacy-preserving PD detection in daily life.

Channel information plays an important role in modern wireless communication systems. Systems that use different frequency bands for uplink and downlink communication often need feedback between devices to exchange band specific channel information. The current state-of-the-art approach proposes a way to predict the channel in the downlink based on that of the observed uplink by identifying variables underlying the uplink channel. In this paper we present a solution that greatly reduces the complexity of this task, and is even applicable for single antenna devices. Our approach uses a neural network trained on a standard channel model to generate coarse estimates for the variables underlying the channel. We then use a simple and efficient single antenna optimization framework to get more accurate variable estimates, which can be used for downlink channel prediction. We implement our approach on software defined radios and compare it to the state-of-the-art through experiments and simulations. Results show that our approach reduces the time complexity by at least an order of magnitude (10x), while maintaining similar prediction quality.

Most existing Augmented Reality (AR) and Mixed Reality (MR) systems are able to understand the 3D geometry of the surroundings but lack the ability to detect and classify complex objects in the real world. Such capabilities can be enabled with deep Convolutional Neural Networks (CNN), but it remains difficult to execute large networks on mobile devices. Offloading object detection to the edge or cloud is also very challenging due to the stringent requirements on high detection accuracy and low end-to-end latency. The long latency of existing offloading techniques can significantly reduce the detection accuracy due to changes in the user's view. To address the problem, we design a system that enables high accuracy object detection for commodity AR/MR system running at 60fps. The system employs low latency offloading techniques, decouples the rendering pipeline from the offloading pipeline, and uses a fast object tracking method to maintain detection accuracy. The result shows that the system can improve the detection accuracy by 20.2%-34.8% for the object detection and human keypoint detection tasks, and only requires 2.24ms latency for object tracking on the AR device. Thus, the system leaves more time and computational resources to render virtual elements for the next frame and enables higher quality AR/MR experiences.

Smart speakers allow users to interact with home appliances using voice commands and are becoming increasingly popular. While voice-based interface is intuitive, it is insufficient in many scenarios, such as in noisy or quiet environments, for users with language barriers, or in applications that require continuous motion tracking. Motion-based control is attractive and complementary to existing voice-based control. However, accurate and reliable room-scale motion tracking poses a significant challenge due to low SNR, interference, and varying mobility. To this end, we develop a novel recurrent neural network (RNN) based system that uses speakers and microphones to realize accurate room-scale tracking. Our system jointly estimates the propagation distance and angle-of-arrival (AoA) of signals reflected by the hand, based on AoA-distance profiles generated by 2D MUSIC. We design a series of techniques to significantly enhance the profile quality under low SNR. We feed the profiles in a recent history to our RNN to estimate the distance and AoA. In this way, we can exploit the temporal structure among consecutive profiles to remove the impact of noise, interference and mobility. Using extensive evaluation, we show our system achieves 1.2--3.7~cm error within 4.5~m range, supports tracking multiple users, and is robust against ambient sound. To our knowledge, this is the first acoustic device-free room-scale tracking system.

The rapid growth in the number of mobile devices, subscriptions and their associated traffic, has served as motivation for several projects focused on improving mobile users' quality of experience (QoE). Few have been as contentious as the Google-initiated Accelerated Mobile Project (AMP), both praised for its seemingly instant mobile web experience and criticized based on concerns about the enforcement of its formats. This paper presents the first characterization of AMP's impact on users' QoE. We do this using a corpus of over 2,100 AMP webpages, and their corresponding non-AMP counterparts, based on trendy-keyword-based searches. We characterized AMP's impact looking at common web QoE metrics, including Page Load Time, Time to First Byte and SpeedIndex (SI). Our results show that AMP significantly improves SI, yielding on average a 60% lower SI than non-AMP pages without accounting for prefetching. Prefetching of AMP pages pushes this advantage even further, with prefetched pages loading over 2,000ms faster than non-prefetched AMP pages. This clear boost may come, however, at a non-negligible cost for users with limited data plans as it incurs an average of over 1.4~MB of additional data downloaded, unbeknownst to users.