Automated media classification techniques like speech processing and face recognition are becoming increasingly commonplace and sophisticated. While such tools can add great value to the public sphere, media searches often process sensitive information, leading to a potential breach of client privacy. Thus, there is great potential for applications involving privacy-preserving searches on public databases like Google Images, Flickr, or ``Wanted Persons" directories put forth by various police agencies. The objective of this thesis is to argue that private media searches masking the client's query from the server are both important and practically feasible. The main contributions include an audio search tool that uses private queries to identify a noisy sound clip from a database without giving the database information about the query. The proposed scheme is shown to have computation and communication costs that are sublinear in database size. An important message of this work is that good private search schemes will typically require special algorithms that are designed for the private domain. To that end, some techniques used in the private audio search tool are generalized to adapt nearest-neighbor searches to the private domain. The resulting private nearest-neighbor algorithm is demonstrated in the context of a privacy-preserving face recognition tool.

关键词：媒体分类技术；语音处理；人脸识别；自动化；搜索；敏感信息；隐私

查看摘要 索取原文[89页]

Using current frameworks developers must meticulously track the proper sequence of touch events from multiple temporally overlapping touch streams using disparate event-handling callbacks. In addition, multiple gestures often begin with the same touch event sequence leading to gesture conflicts in which user input is ambiguous. Thus, developers must perform heavy runtime testing to detect and resolve conflicts. We simplify multitouch gesture creation and management with Proton, a framework that allows developers to declaratively specify a gesture as a regular expression of customizable touch event symbols. Proton provides automatic gesture matching and static analysis of gesture conflicts. We also introduce gesture tablature, a graphical notation that concisely describes the sequencing of interleaved touch events. Finally, we present a user study that indicates that users can read and interpret gesture tablature over four times faster than event-handling pseudocode. Multitouch applications require new design principles and tools for development. This dissertation addresses the challenges of designing gestures and interfaces that benefit from multitouch input and presents tools to help developers build and recognize multitouch gestures. This work serves to facilitate a wider adoption of multitouch interfaces. We conclude with several research directions for continuing the investigation of multitouch input.

关键词：触摸流；跟踪触摸；应用程序设计；多点触控

查看摘要 索取原文[136页]

This is particularly true in urban environments where tall buildings block satellite view for GPS, and distortions in Earth’s magnetic field from power lines adversely affect compass readings.In this thesis, we describe an image based  localization  algorithm  for  estimating  the  pose  of  cell  phones  in  urban  environments.  This  is motivated by the fact that most of today’s cell phones are equipped with cameras whose imagery can be matched against an  image database for  localization purposes.

关键词：移动设备；手机；定位算法；GPS

查看摘要 索取原文[51页]

The National Research Council has highlighted the importance of the changeover to multicore architectures on software performance. They predict strong economic impacts for users in industry as a result of the change. This thesis is part of a larger research project that aims to quantify those effects. The main contribution of this thesis is to build a measure of how well firms can take advantage of multicore chips, which is a key ingredient in addressing this important problem. This measure is built up using unique data on firm software usage from the marketing firm Harte Hanks, and combining it with results from the Berkeley Software Parallelism Survey – a survey of parallelism experts run for this thesis. Alternative formulations for creating this parallelism measure are also discussed and evaluated, with the reasons for choosing the main formulation clearly expressed. Finally, this thesis highlights the limitations of this measure and discusses the other on-going initiatives that are underway to address them. This thesis also makes a supplemental contribution by highlighting a new case study detailing the impact of the changeover to multicore. It also reports additional results from the Berkeley Software Parallelism Survey that may be of interest to those assessing the presence of different types of parallelism or different computational motifs in mainstream software programs.

关键词：软件性能；转换多核；芯片；市场营销；并行度

查看摘要 索取原文[44页]

We present a mobile photo editing application— CrowdBrush, which uses Google Android as front-end platform and Amazon Mechanical Turk as back-end platform. As there are increasingly more images were captured by smart phone these days, providing photo manipulation functions on the mobile phone directly become very important for better user experience.Our system aims at providing more advanced photo manipulation features by uploading a user’s image to our backend server, having workers on the internet help to process the image, and then sending the processed result back to the mobile phone.

关键词：CrowbBrush；图片编辑软件；计算机视觉技术；自动全景

查看摘要 索取原文[23页]

In this thesis, we present two new deployment models for middleboxes which offer new capabilities for middlebox usage as well as new business models for middlebox deployment. Netcalls is an extension to the Internet architecture that allows end host applications to invoke and configure middleboxes in any network their traffic traverses; for example, we present a web server that invokes inter-domain DDoS defense when it detects that it is under attack. APLOMB is a system that allows enterprise networks (as well as individual end hosts) to tunnel their traffic to and from a cloud service that applies middlebox processing to their traffic, avoiding the costly and management-intensive burden of administering middleboxes in a local network. Netcalls and APLOMB allow ISPs and cloud providers (respectively) to monetize their deployment of middleboxes by offering them as a service to third-party clients; all the while presenting new capabilities, in the case of netcalls by enabling application interaction and in the case of APLOMB by providing better scalability and easier management. We discuss both of these proposals and their benefits in detail; we then discuss challenges and opportunities towards their deployment and adoption.

关键词：互联网；云计算；加工服务；体系结构

查看摘要 索取原文[80页]

New architectures require careful examination of security properties in order to assess and contain new threats. In light of this, emerging technologies, such as web APIs, medical devices, and applications on mobile phones, are a new security landscape that has recurring security problems. We develop new techniques to analyze these applications for security vulnerabilities, utilizing techniques including: dynamic symbolic execution, binary analysis and reverse engineering, and wide scale application comparison and classification. We develop Kudzu, a system for symbolic execution of JavaScript, and use it to evaluate a wide variety of JavaScript applications in order to find client-side validation vulnerabilities.

关键词：新兴架构；安全漏洞；新兴技术；web APIs；医疗器械；移动应用程序

查看摘要 索取原文[100页]

We present a system for learning nouns directly from images, using probabilistic predictions generated by visual classifiers as the input to Bayesian word learning, and compare this system to human performance in an automated, large-scale experiment. The system captures a significant proportion of the variance in human responses. Combining the uncertain outputs of the visual classifiers with the ability to identify an appropriate level of abstraction that comes from Bayesian word learning allows the system to outperform alternatives that either cannot deal with visual stimuli or use a more conventional computer vision approach.

关键词：机器学习系统；贝叶斯模型；视觉刺激

查看摘要 索取原文[11页]

Current smartphone platforms provide ways for users to control access to information about their location. For instance, on the iPhone, when an application requests access to location information, the operating system asks the user whether to grant location access to this application. In this paper, we study how users are using these controls. Do iPhone users allow applications to access their location? Do their decisions di er from application to application? Can we predict how a user will respond for a particular application, given their past responses for other applications?

关键词：位置信息；隐私；用户行为；应用程序；智能手机

查看摘要 索取原文[30页]

The non-determinism in concurrent and distributed systems and the unreliability of the hardware environment in which they operate can result in defects that are hard to find and understand. In this thesis, we have developed tools and techniques to augment testing to enable it to quickly find and reproduce important bugs in concurrent and distributed systems.

关键词：分布式计算；程序分析；硬件环境；并发云系统；分布式系统

查看摘要 索取原文[195页]