Search

The proliferation of mobile augmented reality applications and the toolkits to create them have serious implications for user privacy. In this paper, we explore how malicious AR app developers can leverage capabilities offered by commercially available AR libraries, and describe how edge computing can be used to address this privacy problem.

Adversarial examples (AEs) are crafted by adding human-imperceptible perturbations to inputs such that a machine-learning based classifier incorrectly labels them. They have become a severe threat to the trustworthiness of machine learning. While AEs in the image domain have been well studied, audio AEs are less investigated. Recently, multiple techniques are proposed to generate audio AEs, which makes countermeasures against them urgent. Our experiments show that, given an audio AE, the transcription results by Automatic Speech Recognition (ASR) systems differ significantly (that is, poor transferability), as different ASR systems use different architectures, parameters, and training datasets. Based on this fact and inspired by Multiversion Programming, we propose a novel audio AE detection approach MVP-Ears, which utilizes the diverse off-the-shelf ASRs to determine whether an audio is an AE. We build the largest audio AE dataset to our knowledge, and the evaluation shows that the detection accuracy reaches 99.88%. While transferable audio AEs are difficult to generate at this moment, they may become a reality in future. We further adapt the idea above to proactively train the detection system for coping with transferable audio AEs. Thus, the proactive detection system is one giant step ahead of attackers working on transferable AEs.

Smartphone-based augmented reality (AR) applications like the Pokemon Go game are increasingly popular. These applications make extensive use of the smartphone's camera, which may pose a privacy problem as the camera may capture sensitive images or audio information. In this paper, we present PrivacyManager, a framework that allows application developers and domain administrators to control potentially privacy-invading functionality of mobile AR applications when faced with undesirable environmental factors. Using contextual information about a user's environment, such as network signals, ambient light and sound, location, and other inputs, our framework is able to inform subscribing applications of conditions which violate a set of pre-configured constraints. This functionality is available to the subscribing applications with a minimum of processing overhead, including additional performance enhancements to streamline location-awareness with as little resource consumption as possible.

Police body cameras are increasingly being used in many places. While body cameras do promote accountability, there are concerns over the loss of privacy as a result of these cameras. Police departments have tried to balance these two concerns by training officers on what incidents to record, and what not to record. However, the existing approach is error prone. This paper describes an encryption-based approach that protects privacy, while still ensuring relevant videos are always captured. Our preliminary investigations suggest that our approach is feasible on existing mobile hardware.

Recent advances in visual data acquisition and Internet technologies make it convenient and popular to collect and share videos. These activities, however, also raise the issue of privacy invasion. One potential privacy threat is the unauthorized capture and/or sharing of covert videos, which are recorded without the awareness of the subject(s) in the video. Automatic classification of such videos can provide an important basis toward addressing relevant privacy issues. The task is very challenging due to the large intra-class variation and between-class similarity, since there is no limit in the content of a covert video and it may share very similar content with a regular video. The challenge brings troubles when applying existing content-based video analysis methods to covert video classification. In this paper, we propose a novel descriptor, codebook growing pattern (CGP), which is derived from latent Dirichlet allocation (LDA) over optical flows. Given an input video V, we first represent it with a sequence of histograms of optical flow (HOF). After that, these HOFs are fed into LDA to dynamically generate the codebook for V. The CGP descriptor is then defined as the growing codebook sizes in the LDA procedure. CGP fits naturally for covert video representation since (1) optical flows can capture the camera motion that characterizes the covert video acquisition, and (2) CGP by itself is insensitive to video content. To evaluate the proposed approach, we collected a large covert video dataset, the first such dataset to our knowledge, and tested the proposed method on the dataset. The results show clearly the effectiveness of the proposed approach in comparison with other state-of-the-art video classification algorithms.

There is a current interest towards outfitting police officers with body cameras while on patrol. While these systems can provide useful video recordings to serve as evidence, there are concerns that widespread deployment of these cameras will erode our individual privacy. This paper looks at the different guidelines for body cameras to identify the main privacy protection mechanisms, as well as suggest possible enhancements to their security.

This study investigated several smartphone-based applications designed for citizens to record encounters with law enforcement officers. The study aimed to analyze the different applications based on their initialization speed, ease of recording, upload speed and robustness. The study also sought to provide the developers of future applications with insight into what attributes are most important to this type of applications. We found that in general, applications of this kind do not always function consistently. The design features that proved the most beneficial were uploading in a broadcast style, providing ample information regarding the status of an upload, and saving a local copies of videos recorded in the app.

The move towards deploying body cameras for law enforcement personnel makes the security and privacy of these body cameras a pressing problem. Body cameras record police interactions with the public, mainly to provide evidence of potential malicious police actions. However, since these body cameras are deployed and maintained by the police departments themselves, there is lack of trust in the integrity of the footage. Furthermore, the pervasive use of body cameras increases the surveillance of general public, causing a loss of privacy. This paper presents a system that protects the integrity of the body camera videos, as well as protect public privacy. Our approach integrates computer vision techniques onto a resource constrained body camera system, and our evaluation indicates that our approach is feasible to be applied on body cameras.

Mobile surveillance systems like vehicle mounted cameras (i.e. police dash cams) and body cameras play an important role in securing physical spaces. However, these systems can generate too much data that hampers timely analysis. Furthermore, existing infrastructure wireless networks may not have sufficient bandwidth to support real time streaming of videos to backend servers. This paper presents the design of SmartArgos, a control framework to regulate mobile surveillance systems. SmartArgos uses software defined networks and computer vision algorithms to manage multiple mobile surveillance streams. Our experiments indicate that SmartArgos makes video streaming efficient and saves bandwidth compared to traditional video uploading from multiple clients.

Cellular networks will periodically experience instances where the network capacity is insufficient to meet network demand. This paper proposes WiGroup, a group formation algorithm to help cellular networks organize mobile devices into groups to reduce cellular network overhead while limiting disruption to the end user. WiGroup is based on extending device-to-device (D2D) communications, a standard current supported by the cellular network operators, to create device-to-group (D2G) communications. Experimental results indicate that our framework is able to reduce up to 37% of the direct cellular connections. Also, trace-driven results show its ability of consistently reducing direct cellular connections by as high as 28%.

Mobile peer-to-peer applications require devices to network themselves on-the-fly to communicate directly with each another. This paper presents mQual, a framework to help create such networks that meet different application requirements, and is able to adjust the network to ensure that these requirements are met in dynamic environments. Our prototype mQual extends the current WiFi-Direct in Android, and the experimental results suggests that mobile apps built using mQual outperform those built using WiFi-Direct.

For this study the techniques of virtual machine migration are understood and the affects deduplication has on migration are evaluated. The benefits of using deduplication and compression on virtual machines show in the metric of space saved during migrating. Deduplication is computationally expensive so we evaluate how to group virtual machines with similar elements in order to improve migration. From this study, grouping virtual machines based on similar elements improves the overhead from deduplication and compression but estimates which virtual machines are best grouped together.

The IEEE 802.16 standard (WiMAX) is an important next-generation networking technology which promises high-speed network access for both mobile and fixed users. In this paper we present a method to estimate link quality for devices connected to Temple University's WiMAX network as they traverse both the main campus and the city of Philadelphia via foot and motor vehicle. This is accomplished by first measuring receive signal strength indicator (RSSI), carrier to interference plus noise ratio (CINR), and bandwidth. After capturing these values, we then analyze the data to provide an estimation of the actual system rate. We then present an approach to predict future states of link quality both while stationary at Temple and when traversing Philadelphia via bus.

Elasticity is the ability to scale computing resources such as memory on-demand, and is one of the main advantages of utilizing cloud computing services. With the increasing popularity of cloud based storage, it is natural that more deduplication based storage systems will be migrated to the cloud. Existing deduplication systems however, do not adequately take advantage of elasticity. In this paper, we illustrate how to use elasticity to improve deduplication based systems, and propose EAD (elasticity aware deduplication), an indexing algorithm that uses the ability to dynamically increase memory resources to improve overall deduplication performance. Our experimental results indicate that EAD is able to detect more than 98\% of all duplicate data, however only consumes less than 5\% of expected memory space. Meanwhile, it claims four times of deduplication efficiency than the state-of-art sampling technique while costs less than half of the amount of memory.

Mobile social networks are vulnerable to Sybil attacks. By creating a large number of fake identities, malicious users can gain a disproportionately high benefit through a Byzantine fashion. Most social network-based Sybil defenses adopt the assumptions that the honest region is a fast-mixing network. However, more and more evidence shows that some real social networks are not fast-mixing, especially when only strong-trust relations are considered. Moreover, the accuracy of all existing solutions is related to the number of attack edges that the adversary can build. In this paper, for addressing these problems, we propose a local ranking system for estimating trust-level between users. Our scheme has three unique features. First, our system is based on both trust and distrust relations. Second, instead of storing the entire social graph, users carry limited information related to themselves. Last but not least, our system weakens the impacts of attack edges by removing several suspicious edges with high centrality. We validate the effectiveness of our solutions through comprehensive experiments.

The popularity of cloud computing is hampered by privacy concerns of entrusting data to a third party. Applying encryption tries to address the privacy problem by preventing the cloud from learning any information. However, many algorithms and software written to upload or backup data to the cloud were designed by implicitly assuming unencrypted data. In this paper, we conducted experiments to determine the impact of encryption on data uploads to the cloud, and suggest future areas of research.

In this paper, a user-based CPU verification scheme is proposed for cloud cheating detection. In this scheme, a predefined computational task is constructed for the cloud to execute in our cheating detection process. Then we compare the difference of the actual execution time (recorded by the user) and the theoretical execution time, as to determine whether the cloud is cheating or not. A time-lock puzzle is introduced to construct the predefined computational task, so that the predefined computational task is guaranteed to be executed by the cloud. Our cheating detection process has a higher probability of detecting cloud cheating if using a larger predefined computational task, which in turn costs more time. Further analysis shows that, if the total detection time is limited, it is better to detect cloud cheating using small-scale and short-length cheating detecting processes multiple times, as opposed to large-scale and long-length processes a few times. Finally, the feasibility and validity of the proposed scheme is shown in the evaluations.

A crucial component of vehicular network security is to establish a secure wireless channel between any two vehicles. In this paper, we propose a scheme to allow two cars to extract a secret key from RSSI (Received Signal Strength Indicator) values in such a way that nearby cars cannot obtain the same secret. Our solution can be executed in noisy, outdoor vehicular environments. We also propose an online parameter learning mechanism to adapt to different channel conditions. We conduct extensive realworld experiments to validate our solution.

Personal Health Records are starting to get more attention, as they have the potential to improve the quality of care and reduce healthcare costs. In this paper, the concepts behind PHRs and the advantages of using PHRs compared to other methods of medical record keeping are discussed. In particular, platform style PHRs that allow the easy incorporation of third party tools are examined, as they have the potential to allow medical data to be used in novel ways.