INFOCOM 2020 论文列表

Content Delivery Networks (CDNs) are critical for optimizing Internet video delivery. In this paper, we characterize how CDNs serve video content, and the implications for video performance, especially emerging 4K video streaming. Our work is based on measurements of multiple well known video publishers served by top CDNs. Our results show that (i) video chunks in a session often see heterogeneous behavior in terms of whether they hit in the CDN, and which layer they are served from; and (ii) application throughput can vary significantly even across chunks in a session based on where they are served from. The differences while sensitive to client location and CDN can sometimes be significant enough to impact the viability of 4K streaming. We consider the implications for Adaptive bit rate (ABR) algorithms given they rely on throughput prediction, and are agnostic to whether objects hit in the CDN and where. We evaluate the potential benefits of exposing where a video chunk is served from to the client ABR algorithm in the context of the widely studied model predictive control (MPC) algorithm. Emulation experiments show the approach has the potential to reduce prediction inaccuracies, and enhance video streaming performance.

Approximate counters play an important role in many computer domains like network measurement, parallel computing and machine learning because they can reduce the required memory cost. With the emergence of new application needs in these domains like flow counting and parallel measuring, simple Morris counters fail to solve them. Therefore, a more general Morris counter is required. However, there has been a lack of complete theoretical research on the statistical properties of this new approximate counter so far.This paper conducts a deep analysis on general Morris counters and derives the minimum upper bound of the variance. To our best knowledge, this is the first work to thoroughly analyze the statistical properties of general Morris counters in theory. Besides, application scenarios are analyzed, showing that conclusions obtained by our research are effective in testing the performance of approximate counters and guiding system architecture design according to accuracy needs.

In this paper, we consider a joint beam tracking and pattern optimization problem for massive multiple input multiple output (MIMO) systems in which the base station (BS) selects a beamforming codebook and performs adaptive beam tracking taking into account the user mobility. A joint adaptation scheme is developed in a two-phase reinforcement learning framework which utilizes practical signaling and feedback information. In particular, an inner agent adjusts the transmission beam index for a given beamforming codebook based on short-term instantaneous signal-to-noise ratio (SNR) rewards. In addition, an outer agent selects the beamforming codebook based on long-term SNR rewards. Simulation results demonstrate that the proposed online learning outperforms conventional codebook-based beamforming schemes using the same number of feedback information. It is further shown that joint beam tracking and beam pattern adaptation provides a significant SNR gain compared to the beam tracking only schemes, especially as the user mobility increases.

The advancement of voice controllable systems (VC-Ses) has dramatically affected our daily lifestyle and catalyzed the smart home's deployment. Currently, most VCSes exploit automatic speaker verification (ASV) to prevent various voice attacks (e.g., replay attack). In this study, we present VMask, a novel and practical voiceprint mimicry attack that could fool ASV in smart home and inject the malicious voice command disguised as a legitimate user. The key observation behind VMask is that the deep learning models utilized by ASV are vulnerable to the subtle perturbations in the voice input space. To generate these subtle perturbations, VMask leverages the idea of adversarial examples. Then by adding the subtle perturbations to the recordings from an arbitrary speaker, VMask can mislead the ASV into classifying the crafted speech samples, which mirror the former speaker for human, as the targeted victim. Moreover, psychoacoustic masking is employed to manipulate the adversarial perturbations under human perception threshold, thus making victim unaware of ongoing attacks. We validate the effectiveness of VMask by performing comprehensive experiments on both grey box (VGGVox) and black box (Microsoft Azure Speaker Verification) ASVs. Additionally, a real-world case study on Apple HomeKit proves the VMask's practicability on smart home platforms.

Mobile crowdsensing, through which a requester can coordinate a crowd of workers to complete some sensing tasks, has attracted significant attention recently. In this paper, we focus on the unknown worker recruitment problem in mobile crowdsensing, where workers' sensing qualities are unknown a priori. We consider the scenario of recruiting workers to complete some continuous sensing tasks. The whole process is divided into multiple rounds. In each round, every task may be covered by more than one recruited workers, but its completion quality only depends on these workers' maximum sensing quality. Each recruited worker will incur a cost, and each task is attached a weight to indicate its importance. Our objective is to determine a recruiting strategy to maximize the total weighted completion quality under a limited budget. We model such an unknown worker recruitment process as a novel combinatorial multi-armed bandit problem, and propose an extended UCB based worker recruitment algorithm. Moreover, we extend the problem to the case where the workers' costs are also unknown and design the corresponding algorithm. We analyze the regrets of the two proposed algorithms and demonstrate their performance through extensive simulations on real-world traces.

With the advent of the era of big data, deep learning has become a prevalent building block in a variety of machine learning or data mining tasks, such as signal processing, network modeling and traffic analysis, to name a few. The massive user data crowdsourced plays a crucial role in the success of deep learning models. However, it has been shown that user data may be inferred from trained neural models and thereby exposed to potential adversaries, which raises information security and privacy concerns. To address this issue, recent studies leverage the technique of differential privacy to design private-preserving deep learning algorithms. Albeit successful at privacy protection, differential privacy degrades the performance of neural models. In this paper, we develop ADADP, an adaptive and fast convergent learning algorithm with a provable privacy guarantee. ADADP significantly reduces the privacy cost by improving the convergence speed with an adaptive learning rate and mitigates the negative effect of differential privacy upon the model accuracy by introducing adaptive noise. The performance of ADADP is evaluated on real-world datasets. Experiment results show that it outperforms state-of-the-art differentially private approaches in terms of both privacy cost and model accuracy.

The Age-of-Information (AoI) is a newly introduced metric for capturing information updating timeliness, as opposed to the network throughput, which is a conventional performance metric to measure the network transmission speed and robustness as a whole. While considerable work has addressed either optimal AoI or throughput individually, the inherent relationships between the two performance metrics are yet to be explored, especially in multi-hop networks. In this paper, we explore their relationships in multi-hop networks for the very first time, particularly focusing on the impacts of flexible routes on the two metrics. By developing a rigorous mathematical model with interference, channel allocation, link scheduling, and routing path selection taken into consideration, we build the interrelation between AoI and throughput in multi-hop networks. A multi-criteria optimization problem is formulated with the goal of simultaneously minimizing AoI and maximizing network throughput. To solve this problem, we resort to a novel approach by transforming the multi-criteria problem into a single objective one so as to find the weakly Pareto-optimal points iteratively, thereby allowing us to screen all Pareto-optimal points for the solution. A new algorithm based on the piece-wise linearization technique is then developed to closely linearize the non-linear terms in the single objective problem via their linear approximation segments to make it solvable. We formally prove that our algorithms can find all Pareto-optimal points in a finite number of iterations. From simulation results, we identify the tradeoff points of the optimal AoI and throughput, demonstrating that one performance metric improves at the expense of degrading the other, with the routing path found as one of the key factors in determining such a tradeoff.

Monitoring the performance of a large network is very costly. Instead, a subset of paths or time intervals of the network can be measured while inferring the remaining network data by leveraging their spatiotemporal correlations. The quality of missing data recovery highly relies on the inference algorithms. Tensor completion has attracted some recent attentions with its capability of exploiting the multi-dimensional data structure for more accurate missing data inference. However, current tensor completion algorithms only model the three-order interaction of data features through the inner product, which is insufficient to capture the high-order, nonlinear correlations across different feature dimensions. In this paper, we propose a novel Neural Tensor Completion (NTC) scheme to effectively model three-order interaction among data features with the outer product and build a 3D interaction map. Based on which, we apply 3D convolution to learn features of high-order interaction from the local range to the global range. We demonstrate this will lead to good learning ability. We conduct extensive experiments on two real-world network monitoring datasets, Abilene and WS-DREAM, to demonstrate that NTC can significantly reduce the error in missing data recovery. When the sampling ratio is low at 1%, the recovery error ratios on the testing data are around 0.05 (Abilene) and 0.13 (WS-DREAM) when using NTC, but are 0.99 (Abilene) and 0.99 (WS-DREAM) using the best current tensor completion algorithms, which are 21 times and 8 times larger.

Although existing surveillance cameras can identify people, their utility is limited by the unavailability of any direct camera-to-human communication. This paper proposes a real-time end-to-end system to solve the problem of digitally associating people in a camera view with their smartphones, without knowing the phones' IP/MAC addresses. The key idea is using a person's unique "context features", extracted from videos, as its sole address. The context address consists of: motion features, e.g. walking velocity; and ambience features, e.g. magnetic trend and Wi-Fi signal strengths. Once receiving a broadcast packet from the camera, a user's phone accepts it only if its context address matches the phone's sensor data. We highlight three novel components in our system: (1) definition of discriminative and noise-robust ambience features; (2) effortless ambient sensing map generation; (3) a context feature selection algorithm to dynamically choose lightweight yet effective features which are encoded into a fixed-length header. Real-world and simulated experiments are conducted for different applications. Our system achieves a sending ratio of 98.5%, an acceptance precision of 93.4%, and a recall of 98.3% with ten people. We believe this is a step towards direct camera-to-human communication and will become a generic underlay to various practical applications.

Recent studies have suggested that the BitTorrent's rarest-first protocol, owing to its work-conserving nature, can become unstable in the presence of non-persistent users. Consequently, in any stable protocol, many peers are at some point endogenously forced to hold off their file-download activity. In this work, we propose a tunable piece-selection policy that minimizes this (undesirable) requisite by combining the (work-conserving) rarest-first protocol with only an appropriate share of the (non-work conserving) mode-suppression protocol. We refer to this policy as "Rarest-First with Probabilistic Mode-Suppression" or simply RFwPMS. We study RFwPMS under a stochastic model of the BitTorrent network that is general enough to capture multiple swarms of non-persistent users - each swarm having its own altruistic preferences that may or may not overlap with those of other swarms. Using a Lyapunov drift analysis, we show that RFwPMS is provably stable for all kinds of inter-swarm behaviors, and that the use of rarest-first instead of random-selection is indeed more justified. Our numerical results suggest that RFwPMS is scalable in the general multi-swarm setting and offers better performance than the existing stabilizing schemes like mode-suppression.

Cellular network operators deploy base stations with a high density to ensure radio signal coverage for 4G/5G networks. While users enjoy the high-speed connection provided by cellular networks, an adversary could exploit the dense cellular deployment to detect nearby human movements and even recognize keystroke movements of a victim by passively listening to the CRS broadcast from base stations. To demonstrate this, we develop SpiderMon, the first attempt to perform passive continuous keystroke monitoring using the signal transmitted by commercial cellular base stations. Our experimental results show that SpiderMon can detect keystroke movements at a distance of 15 meters and can recover a 6-digits PIN input with a success rate of more than 51% within ten trials when the victim is behind the wall.

We explore the feasibility of Multiple-Input-Multiple-Output (MIMO) communication through vibrations over human skin. Using off-the-shelf motors and piezo transducers as vibration transmitters and receivers, respectively, we build a 2x2 MIMO testbed to collect and analyze vibration signals from real subjects. Our analysis reveals that there exist multiple independent vibration channels between a pair of transmitter and receiver, confirming the feasibility of MIMO. Unfortunately, the slow ramping of mechanical motors and rapidly changing skin channels make it impractical for conventional channel sounding based channel state information (CSI) acquisition, which is critical for achieving MIMO capacity gains. To solve this problem, we propose Skin-MIMO, a deep learning based CSI acquisition technique to accurately predict CSI entirely based on inertial sensor (accelerometer and gyroscope) measurements at the transmitter, thus obviating the need for channel sounding. Based on experimental vibration data, we show that Skin-MIMO can improve MIMO capacity by a factor of 2.3 compared to Single-Input-Single-Output (SISO) or open-loop MIMO, which do not have access to CSI. A surprising finding is that gyroscope, which measures the angular velocity, is found to be superior in predicting skin vibrations than accelerometer, which measures linear acceleration and used widely in previous research for vibration communications over solid objects.

Research on Cross-technology communication (CTC) has made rapid progress in recent years, but how to estimate the quality of a CTC link remains an open and challenging problem. Through our observation and study, we find that none of the existing approaches can be applied to estimate the link quality of CTC. Built upon the physical-level emulation, transmission over a CTC link is jointly affected by two factors: the emulation error and the channel distortion. We in this paper propose a new link metric called C-LQI and a joint link model that simultaneously takes into account the emulation error and the channel distortion in the process of CTC. We further design a light-weight link estimation approach to estimate C-LQI and in turn the PRR over the CTC link. We implement C-LQI and compare it with two representative link estimation approaches. The results demonstrate that C-LQI reduces the relative error of link estimation respectively by 46% and 53% and saves the communication cost by 90%.

Millimeter-wave systems not only provide high data rates and low latency, but the very large bandwidth also allows for highly accurate environment sensing. Such properties are extremely useful for smart factory scenarios. At the same time, reusing existing communication links for passive object localization is significantly more challenging than radar-based approaches due to the sparsity of the millimeter-wave multi-path environment and the weakness of the reflected paths compared to the line-of-sight path. In this paper, we explore the passive object localization accuracy that can be achieved with IEEE 802.11ad devices. We use commercial Access Points (APs) whereas the station design is based on a full-bandwidth 802.11ad compatible FPGA-based platform with a phased antenna array. The stations exploit the preamble of the beam training packets of the APs to obtain Channel Impulse Response (CIR) measurements for all antenna patterns. With this, we determine distance and angle information for the different multi-path components in the environment to passively localize a mobile object. We evaluate our system with multiple APs and a moving robot with a metallic surface. Our system operates in real-time and achieves 6.5cm mean error accuracy and sub-meter accuracy in 100% of the cases.

In large-scale online service system, to enhance the quality of services, engineers need to collect various monitoring data and write many rules to trigger alerts. However, the number of alerts is way more than what on-call engineers can properly investigate. Thus, in practice, alerts are classified into several priority levels using manual rules, and on-call engineers primarily focus on handling the alerts with the highest priority level (i.e., severe alerts). Unfortunately, due to the complex and dynamic nature of the online services, this rule-based approach results in missed severe alerts or wasted troubleshooting time on non-severe alerts. In this paper, we propose AlertRank, an automatic and adaptive framework for identifying severe alerts. Specifically, AlertRank extracts a set of powerful and interpretable features (textual and temporal alert features, univariate and multivariate anomaly features for monitoring metrics), adopts XGBoost ranking algorithm to identify the severe alerts out of all incoming alerts, and uses novel methods to obtain labels for both training and testing. Experiments on the datasets from a top global commercial bank demonstrate that AlertRank is effective and achieves the F1-score of 0.89 on average, outperforming all baselines. The feedback from practice shows AlertRank can significantly save the manual efforts for on-call engineers.

Today's explosively growing Internet video traffics and viewers' ever-increasing quality of experience (QoE) demands for video streaming bring tremendous pressures to the backbone network. As a new network paradigm, mobile edge caching provides a promising alternative by pushing video content closer at the network edge rather than the remote CDN servers so as to reduce both content access latency and redundant network traffic. However, our large-scale trace analysis shows that different from CDN based caching, edge caching environment is much more complicated with massively dynamic and diverse request patterns, which renders that existing rule-based and model-based caching solutions may not well fit such complicated edge environments. Moreover, although cooperative caching has been proposed to better afford limited storage on each individual edge server, our trace analysis also shows that the request similarity among neighboring edges can be highly dynamic and diverse, which is drastically different from CDN based caching environment, and can easily compromise the benefits from traditional cooperative caching mostly designed based on CDN environment. In this paper, we propose MacoCache, an intelligent edge caching framework that is carefully designed to afford the massively diversified and distributed caching environment to minimize both content access latency and traffic cost. Specifically, MacoCache leverages a multi-agent deep reinforcement learning (MADRL) based solution, where each edge is able to adaptively learn its own best policy in conjunction with other edges for intelligent caching. The real trace-driven evaluation further demonstrates that MacoCache is able to reduce an average of 21% latency and 26% cost compared with the state-of-the-art caching solution.

We focus on policy-aware data centers (PADCs), wherein virtual machine (VM) traffic traverses a sequence of middleboxes (MBs) for security and performance purposes, and propose two new VM placement and migration problems. We first study PAL: policy-aware virtual machine placement. Given a PADC with a data center policy that communicating VM pairs must satisfy, the goal of PAL is to place the VMs into the PADC to minimize their total communication cost. Due to dynamic traffic loads in PADCs, however, above VM placement may no longer be optimal after some time. We thus study PAM: policy-aware virtual machine migration. Given an existing VM placement in the PADC and dynamic traffic rates among communicating VMs, PAM migrates VMs in order to minimize the total cost of migration and communication of the VM pairs. We design optimal, approximation, and heuristic policyaware VM placement and migration algorithms. Our experiments show that i) VM migration is an effective technique, reducing total communication cost of VM pairs by 25%, ii) our PAL algorithms outperform state-of-the-art VM placement algorithm that is oblivious to data center policies by 40-50%, and iii) our PAM algorithms outperform the only existing policy-aware VM migration scheme by 30%.

Passive human localization and tracking using RF signals have been studied for over a decade. Most of the existing solutions, however, can only track a single moving subject due to the coarse multipath resolvability limited by bandwidth and antenna number. In this paper, we break down the limitations by leveraging the emerging 60GHz millimeter-wave radios. We present mmTrack, the first system that passively localizes and tracks multiple users simultaneously using a single commodity 60GHz radio. The design of mmTrack consists of three key components. First, we significantly improve the spatial resolution, limited by the small aperture of the compact 60GHz array, by performing digital beamforming over all receive antennas. Second, we propose a novel multi-target detection approach that tackles the near-far-effect and measurement noise. Finally, we devise a robust clustering technique to accurately recognize multiple targets and estimate the respective locations, from which their individual trajectories are further derived by a continuous tracking algorithm. We implement mmTrack on a commodity 802.11ad device and evaluate it in indoor environments. Our experiments demonstrate that mmTrack detects and counts multiple users precisely with an error ≤ 1 person for 97.8% of the time and achieves a respective median location error of 9.9 cm and 19.7 cm for dynamic and static targets.

Geographically distributed cloud is a promising technique to achieve high performance for service providers. For inter-datacenter transfers, deadline guarantee and fairness are the two most important requirements. On the one hand, to ensure more transfers finish before their deadlines, preemptive scheduling policies are widely used, leading to the transfer starvation problem and is hence unfair. On the other hand, to ensure fairness, inter-datacenter bandwidth is fairly shared among transfers with per-flow bandwidth allocation, which leads to deadline missing problem. A mechanism that achieves these two seemingly conflicting objectives simultaneously is still missing. In this paper, we propose TINA to schedule network transfers fairly while providing deadline guarantees. TINA allows each transfer to compete freely with each other for bandwidth. More specifically, each transfer is assigned a probability to indicate whether to transmit or not. We formulate the competition among the transfers as an El Farol game while keeping the traffic load under a threshold to avoid congestion. We then prove that the Nash Equilibrium is the optimal strategy and propose a light-weight algorithm to derive it. Finally, both simulations and testbed experiments results show that TINA achieves superior performance than state-of-art methods in terms of fairness and deadline guarantee rate.