INFOCOM 2021 论文列表

Data holders, such as mobile apps, hospitals and banks, are capable of training machine learning (ML) models and enjoy many intelligence services. To benefit more individuals lacking data and models, a convenient approach is needed which enables the trained models from various sources for prediction serving, but it has yet to truly take off considering three issues: (i) incentivizing prediction truthfulness; (ii) boosting prediction accuracy; (iii) protecting model privacy.We design FedServing, a federated prediction serving framework, achieving the three issues. First, we customize an incentive mechanism based on Bayesian game theory which ensures that joining providers at a Bayesian Nash Equilibrium will provide truthful (not meaningless) predictions. Second, working jointly with the incentive mechanism, we employ truth discovery algorithms to aggregate truthful but possibly inaccurate predictions for boosting prediction accuracy. Third, providers can locally deploy their models and their predictions are securely aggregated inside TEEs. Attractively, our design supports popular prediction formats, including top-1 label, ranked labels and posterior probability. Besides, blockchain is employed as a complementary component to enforce exchange fairness. By conducting extensive experiments, we validate the expected properties of our design. We also empirically demonstrate that FedServing reduces the risk of certain membership inference attack.

Website fingerprinting (WFP) could infer which websites a user is accessing via an encrypted proxy by passively inspecting the traffic between the user and the proxy. The key to WFP is designing a classifier capable of distinguishing traffic characteristics of accessing different websites. However, when deployed in real-life networks, a well-trained classifier may face a significant obstacle of training-testing asymmetry, which fundamentally limits its practicability. Specifically, although pure traffic samples can be collected in a controlled (clean) testbed for training, the classifier may fail to extract such pure traffic samples as its input from raw complicated traffic for testing. In this paper, we are interested in encrypted proxies that relay connections between the user and the proxy individually (e.g., Shadowsocks), and design a context-aware system using built-in spatial-temporal flow correlation to address the obstacle. Extensive experiments demonstrate that our system does not only enable WFP against a popular type of encrypted proxies practical, but also achieves better performance than ideally training/testing pure samples.

Federated learning (FL) enables participants to collaboratively construct a global machine learning model without sharing their local training data to the remote server. In FL systems, the selection of training samples has a significant impact on model performances, e.g., selecting participants whose datasets have erroneous samples, skewed categorical distributions, and low content diversity would result in low accuracy and unstable models. In this work, we aim to solve the exigent optimization problem that selects a collection of high-quality training samples for a given FL task under a monetary budget in a privacy-preserving way, which is extremely challenging without visibility to participants' local data and training process. We provide a systematic analysis of important data related factors affecting the model performance and propose a holistic design to privately and efficiently select high-quality data samples considering all these factors. We verify the merits of our proposed solution with extensive experiments on a real AIoT system with 50 clients, including 20 edge computers, 20 laptops, and 10 desktops. The experimental results validates that our solution achieves accurate and efficient selection of high-quality data samples, and consequently an FL model with a faster convergence speed and higher accuracy than that achieved by existing solutions.

The conventional federated learning (FedL) architecture distributes machine learning (ML) across worker devices by having them train local models that are periodically aggregated by a server. FedL ignores two important characteristics of contemporary wireless networks, however: (i) the network may contain heterogeneous communication/computation resources, while (ii) there may be significant overlaps in devices' local data distributions. In this work, we develop a novel optimization methodology that jointly accounts for these factors via intelligent device sampling complemented by device-to-device (D2D) offloading. Our optimization aims to select the best combination of sampled nodes and data offloading configuration to maximize FedL training accuracy subject to realistic constraints on the network topology and device capabilities. Theoretical analysis of the D2D offloading subproblem leads to new FedL convergence bounds and an efficient sequential convex optimizer. Using this result, we develop a sampling methodology based on graph convolutional networks (GCNs) which learns the relationship between network attributes, sampled nodes, and resulting offloading that maximizes FedL accuracy. Through evaluation on real-world datasets and network measurements from our IoT testbed, we find that our methodology while sampling less than 5% of all devices outperforms conventional FedL substantially both in terms of trained model accuracy and required resource utilization.

Kubernetes (k8s) has the potential to merge the distributed edge and the cloud but lacks a scheduling framework specifically for edge-cloud systems. Besides, the hierarchical distribution of heterogeneous resources and the complex dependencies among requests and resources make the modeling and scheduling of k8s-oriented edge-cloud systems particularly sophisticated. In this paper, we introduce KaiS, a learning-based scheduling framework for such edge-cloud systems to improve the long-term throughput rate of request processing. First, we design a coordinated multi-agent actor-critic algorithm to cater to decentralized request dispatch and dynamic dispatch spaces within the edge cluster. Second, for diverse system scales and structures, we use graph neural networks to embed system state information, and combine the embedding results with multiple policy networks to reduce the orchestration dimensionality by stepwise scheduling. Finally, we adopt a two-time-scale scheduling mechanism to harmonize request dispatch and service orchestration, and present the implementation design of deploying the above algorithms compatible with native k8s components. Experiments using real workload traces show that KaiS can successfully learn appropriate scheduling policies, irrespective of request arrival patterns and system scales. Moreover, KaiS can enhance the average system throughput rate by 14.3% while reducing scheduling cost by 34.7% compared to baselines.

Regulators in the US and Europe have stepped up their efforts to open the 6 GHz bands for unlicensed access. The two unlicensed technologies likely to operate and coexist in these bands are Wi-Fi 6E and 5G New Radio Unlicensed (NR-U). The greenfield 6 GHz bands allow us to take a fresh look at the coexistence between Wi-Fi and 3GPP-based unlicensed technologies. In this paper, using tools from stochastic geometry, we study the impact of Multi User Orthogonal Frequency Division Multiple Access, i.e., MU OFDMA-a feature introduced in 802.11ax-on this coexistence issue. Our results reveal that by disabling the use of the legacy contention mechanism (and allowing only MU OFDMA) for uplink access in Wi-Fi 6E, the performance of both NR-U networks and uplink Wi-Fi 6E can be improved. This is indeed feasible in the 6 GHz bands, where there are no operational Wi-Fi or NR-U users. In so doing, we also highlight the importance of accurate channel sensing at the entity that schedules uplink transmissions in Wi-Fi 6E and NR-U. If the channel is incorrectly detected as idle, factors that improve the uplink performance of one technology contribute negatively to the performance of the other technology.

Recently, vehicular crowd sensing (VCS) that leverages sensor-equipped urban vehicles to collect city-scale sensory data has emerged as a promising paradigm for urban sensing. Nowadays, a wide spectrum of VCS tasks are carried out by for-hire vehicles (FHVs) due to various hardware and software constraints that are difficult for private vehicles to satisfy. However, such FHV-enabled VCS systems face a fundamental yet unsolved problem of striking a balance between the order-serving and sensing outcomes. To address this problem, we propose a novel graph convolutional cooperative multi-agent reinforcement learning (GCC-MARL) framework, which helps FHVs make distributed routing decisions that cooperatively optimize the system-wide global objective. Specifically, GCC-MARL meticulously assigns credits to agents in the training process to effectively stimulate cooperation, represents agents' actions by a carefully chosen statistics to cope with the variable agent scales, and integrates graph convolution to capture useful spatial features from complex large-scale urban road networks. We conduct extensive experiments with a real-world dataset collected in Shenzhen, China, containing around 1 million trajectories and 50 thousand orders of 553 taxis per-day from June 1st to 30th, 2017. Our experiment results show that GCC-MARL outperforms state-of-the-art baseline methods in order-serving revenue, as well as sensing coverage and quality.

Wireless home alarm systems are being widely deployed, but their security has not been well studied. Existing attacks on wireless home alarm systems exploit the vulnerabilities of networking protocols while neglecting the problems arising from the physical component of IoT devices. In this paper, we present new event-eliminating and event-spoofing attacks on commercial wireless home alarm systems by interfering with the reed switch in almost all COTS alarm sensors. In both attacks, the external adversary uses his own magnet to control the state of the reed switch in order to either eliminate legitimate alarms or spoof false alarms. We also present a new battery-depletion attack with programmable electromagnets to deplete the alarm sensor's battery quickly and stealthily in hours which is expected to last a few years. The efficacy of our attacks is confirmed by detailed experiments on a representative Ring alarm system.

Resource Public Key Infrastructure (RPKI) is vital to the security of inter-domain routing. However, RPKI enables Regional Internet Registries (RIRs) to unilaterally takedown IP prefixes - indeed, such attacks have been launched by nation-state adversaries. The threat of IP prefix takedowns is one of the factors hindering RPKI adoption.In this work, we propose the first distributed RPKI system, based on threshold signatures, that requires the coordination of a number of RIRs to make changes to RPKI objects; hence, preventing unilateral prefix takedown. We perform extensive evaluations using our implementation demonstrating the practicality of our solution. Furthermore, we show that our system is scalable and remains efficient even when RPKI is widely deployed.

Current web tracking practices pose a constant threat to the privacy of Internet users. As a result, the research community has recently proposed different tools to combat well-known tracking methods. However, the early detection of new, previously unseen tracking systems is still an open research problem. In this paper, we present TrackSign, a novel approach to discover new web tracking methods. The main idea behind TrackSign is the use of code fingerprinting to identify common pieces of code shared across multiple domains. To detect tracking fingerprints, TrackSign builds a novel 3-mode network graph that captures the relationship between fingerprints, resources and domains. We evaluated TrackSign with the top-100K most popular Internet domains, including almost 1M web resources from more than 5M HTTP requests. Our results show that our method can detect new web tracking resources with high precision (over 92%). TrackSign was able to detect 30K new trackers, more than 10K new tracking resources and 270K new tracking URLs, not yet detected by most popular blacklists. Finally, we also validate the effectiveness of TrackSign with more than 20 years of historical data from the Internet Archive.

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

Today's mmWave WLANs can realize simultaneous multi-user multi-stream transmission solely on the downlink. In this paper, we present Uplink Multi-user Beamforming on single RF chain AP (UMBRA), a novel framework for supporting multi-stream multi-user uplink transmissions via a single RF chain. We design multi-user overlayed constellations and multi-user receiver mechanisms to enable concurrent time-triggered uplink multi-user transmissions received on a single RF chain AP. We devise exemplary beam selection policies to jointly adapt beams at users and the AP for targeting aggregate rate maximization without increasing training requirements compared to single-user systems. We implement the key components of UMBRA using a programmable WLAN testbed using software-defined radios and commercial 60-GHz transceivers and collect over-the-air measurements using phased-array antennas and horn antennas with varying beamwidth. We find that in comparison to single-user transmissions, UMBRA achieves more than 1.45× improvement in aggregate rate regardless of the choice of the user group, geometric separation, and receiver beamwidth.

Web performance anomalies (e.g. time periods when metrics like page load time are abnormally high) have significant impact on user experience and revenues of web service providers. Existing methods to automatically detect web performance anomalies focus on popular websites (e.g. with tens of thousands of visits per minute). Across a wider diversity of websites, however, the number of visits per hour varies enormously, and some sites will only have few visits per hour. Low rates of visits create measurement gaps and noise that prevent the use of existing methods. This paper develops WMF, a web performance anomaly detection method applicable across a range of websites with highly variable measurement volume. To demonstrate our method, we leverage data from a website monitoring company, which allows us to leverage cross-site measurements. WMF uses matrix factorization to mine patterns that emerge from a subset of the websites to "fill in" missing data on other websites. Our validation using both a controlled website and synthetic anomalies shows that WMF's F1-score is more than double that of the state-of-the-art method. We then apply WMF to three months of web performance measurements to shed light on performance anomalies across a variety of 125 small to medium websites.

Interactions between the intra- and inter-domain routing protocols received little attention despite playing an important role in forwarding transit traffic. More precisely, by default, IGP distances are taken into account by BGP to select the closest exit gateway for the transit traffic (hot-potato routing). Upon an IGP update, the new best gateway may change and should be updated through the (full) re-convergence of BGP, causing superfluous BGP processing and updates in many cases. We propose OPTIC (Optimal Protection Technique for Inter-intra domain Convergence), an efficient way to assemble both protocols without losing the hot-potato property. OPTIC pre-computes sets of gateways (BGP next-hops) shared by groups of prefixes. Such sets are guaranteed to contain the post-convergence gateway after any single IGP event for the grouped prefixes. The new optimal exits can be found through a single walk-through of each set, allowing the transit traffic to benefit from optimal BGP routes almost as soon as the IGP converges. Compared to vanilla BGP, OPTIC's structures allow it to consider a reduced number of entries: this number can be reduced by 99% for stub networks. The update of OPTIC's structures, which is not required as long as border routers remain at least bi-connected, scales linearly in time with its number of groups.

Mobile edge computing (MEC) has been an effective paradigm to support real-time computation-intensive vehicular applications. However, due to highly dynamic vehicular topology, these existing centralized-based or distributed-based scheduling algorithms requiring high communication overhead, are not suitable for task offloading in vehicular networks. Therefore, we investigate a novel service scenario of MEC-based vehicular crowdsourcing, where each MEC server is an independent agent and responsible for making scheduling of processing traffic data sensed by crowdsourcing vehicles. On this basis, we formulate a data-driven task offloading problem by jointly optimizing offloading decision and bandwidth/computation resource allocation, and renting cost of heterogeneous servers, such as powerful vehicles, MEC servers and cloud, which is a mixed-integer programming problem and NP-hard. To reduce high time-complexity, we propose the solution in two stages. First, we design an asynchronous deep Q-learning to determine offloading decision, which achieves fast convergence by training the local DQN model at each agent in parallel and uploading for global model update asynchronously. Second, we decompose the remaining resource allocation problem into several independent subproblems and derive optimal analytic formula based on convex theory. Lastly, we build a simulation model and conduct comprehensive simulation, which demonstrates the superiority of the proposed algorithm.

A modern network needs to conduct a diverse set of tasks, and the existing approaches focus on developing specific tools for specific tasks, resulting in increasing complexity and lacking reusability. In this paper, we propose Flow Algebra as a unifying, easy-to-use framework to accomplish a large set of network management tasks. Based on the observation that relational databases based on relational algebra are well understood and widely used as a unifying framework for data management, we develop flow algebra based on relational algebra. On the other hand, flow tables, which are the fundamental data specifying the state of a network, cannot be stored in traditional relations, because of fundamental features such as wildcard and priorities. We define flow algebra based on novel, generalized relational operations that use equivalency to achieve efficient, unifying data store, query, and manipulation of both flow tables and traditional relations. We realize flow algebra with FlowDB and demonstrate its ease of use on diverse tasks. We further demonstrate that generality and ease-of-use do not need to come with a performance penalty. For example, for the well-studied network verification task, our system outperforms two state-of-the-art network verification engines, NoD and HSA, in their targeted domain, by 55x.

The rapid uptake of digital services and Internet of Things (IoT) technology gives rise to unprecedented numbers and diversification of cyber attacks, with which commonly-used rule-based Network Intrusion Detection Systems (NIDSs) are struggling to cope. Therefore, Artificial Intelligence (AI) is being exploited as second line of defense, since this methodology helps in extracting non-obvious patterns from network traffic and subsequently in detecting more confidently new types of threats. Cybersecurity is however an arms race and intelligent solutions face renewed challenges as attacks evolve while network traffic volumes surge. In this paper, we propose Adaptive Clustering-based Intrusion Detection (Acid), a novel approach to malicious traffic classification and a valid candidate for deployment at the network edge. Acid addresses the critical challenge of sensitivity to subtle changes in traffic features, which routinely leads to misclassification. We circumvent this problem by relying on low-dimensional embeddings learned with a lightweight neural model comprising multiple kernel networks that we introduce, which optimally separates samples of different classes. We empirically evaluate our approach with both synthetic and three intrusion detection datasets spanning 20 years, and demonstrate Acid consistently attains 100% accuracy and F1-score, and 0% false alarm rate, thereby significantly outperforming state-of-the-art clustering methods and NIDSs.

Human-machine interface based on tongue-jaw movements has recently become one of the major technological trends. However, existing schemes have several limitations, such as requiring dedicated hardware and are usually uncomfortable to wear. This paper presents CanalScan, a nonintrusive system for tongue-jaw movement recognition using only commodity speaker and microphone mounted on ubiquitous off-the-shelf devices (e.g., smartphones). The basic idea is to send an acoustic signal, then captures its reflections and derive unique patterns of ear canal deformation caused by tongue-jaw movements. A dynamic segmentation method with Support Vector Domain Description is used to segment tongue-jaw movements. To combat sensor position-sensitive deficiency and ear-canal-shape-sensitive deficiency in multi-path reflections, we first design algorithms to assist users in adjusting the acoustic sensors to the same valid zone. Then we propose a data transformation mechanism to reduce the impacts of diversities in ear canal shapes and relative positions between sensors and the ear canal. CanalScan explores twelve unique and consistent features and applies a Random Forest classifier to distinguish tongue-jaw movements. Extensive experiments with twenty participants demonstrate that CanalScan achieves promising recognition for six tongue-jaw movements, is robust against various usage scenarios, and can be generalized to new users without retraining and adaptation.

Crowdsourcing is a widely adopted way for various labor-intensive tasks. One of the core problems in crowdsourcing systems is how to assign tasks to most suitable workers for better results, which heavily relies on the accurate profiling of each worker's reliability for different topics of tasks. Many previous work have studied worker reliability for either explicit topics represented by task descriptions or latent topics for categorical tasks. In this work, we aim to accurately estimate more fine-grained worker reliability for latent topics in numerical tasks, so as to further improve the result quality. We propose a bayesian probabilistic model named Gaussian Latent Topic Model(GLTM) to mine the latent topics of numerical tasks based on workers' behaviors and to estimate workers' topic-level reliability. By utilizing the GLTM, we propose a truth inference algorithm named TI-GLTM to accurately infer the tasks' truth and topics simultaneously and dynamically update workers' topic-level reliability. We also design an online task assignment mechanism called MRA-GLTM, which assigns appropriate tasks to workers with the Maximum Reduced Ambiguity principle. The experiment results show our algorithms can achieve significantly lower MAE and MSE than that of the state-of-the-art approaches.