Many data analysis tasks rely on the abstraction of a graph to represent relations between entities, with attributes on the nodes and edges. Since the relationships encoded are often sensitive, we seek effective ways to release representative graphs which nevertheless protect the privacy of the data subjects. Prior work on this topic has focused primarily on the graph structure in isolation, and has not provided ways to handle richer graphs with correlated attributes.

Mobile platforms, such as Android, warn users about the permissions an app requests and trust that the user will make the correct decision about whether or not to install the app. Unfortunately many users either ignore the warning or fail to understand the permissions and the risks they imply. As a step toward developing an indicator of risk that decomposes risk into several categories, or dimensions, we conducted two studies designed to assess the dimensions of risk deemed most important by experts and novices. In Study 1, semi-structured interviews were conducted with 19 security experts, who also performed a card sorting task in which they categorized permissions. The experts identified three major risk dimensions in the interviews (personal information privacy, monetary risk, and device availability/stability), and a forth dimension (data integrity) in the card sorting task. In Study 2, 350 typical Android users, recruited via Amazon Mechanical Turk, filled out a questionnaire in which they (a) answered questions concerning their mobile device usage, (b) rated how often they considered each of several types of information when installing apps, (c) indicated what they considered to be the biggest risk associated with installing an app on their mobile device, and (d) rated their concerns with regard to specific risk types and about apps having access to specific types of information. In general, the typical users' concerns were similar to those of the security experts. The results of the studies suggest that risk information should be organized into several risk types that can be better understood by users and that a mid-level risk summary should incorporate the dimensions of personal information privacy, monetary risk, device availability/stability risk and data integrity risk.

Differential privacy is widely accepted as a powerful framework for providing strong, formal privacy guarantees for aggregate data analysis. A limitation of the model is that the same level of privacy protection is afforded for all individuals. However, it is common that the data subjects have quite different expectations regarding the acceptable level of privacy for their data. Consequently, differential privacy may lead to insufficient privacy protection for some users, while over-protecting others. We argue that by accepting that not all users require the same level of privacy, a higher level of utility can often be attained by not providing excess privacy to those who do not want it. We propose a new privacy definition called personalized differential privacy (PDP), a generalization of differential privacy in which users specify a personal privacy requirement for their data. We then introduce several novel mechanisms for achieving PDP. Our primary mechanism is a general one that automatically converts any existing differentially private algorithm into one that satisfies PDP. We also present a more direct approach for achieving PDP, inspired by the well-known exponential mechanism. We demonstrate our framework through extensive experiments on real and synthetic data.

We consider the problem of producing item recommenda-tions that are personalized based on a user’s social network, while simultaneously preventing the disclosure of sensitive user-item preferences (e.g., product purchases, ad clicks, web browsing history, etc.). Our main contribution is a privacy-preserving framework for a class of social recommendation algorithms that provides strong, formal privacy guarantees under the model of differential privacy. Existing mechanisms for achieving differential privacy lead to an unacceptable loss of utility when applied to the social recommendation prob-lem. To address this, the proposed framework incorporates a clustering procedure that groups users according to the natural community structure of the social network and sig-nificantly reduces the amount of noise required to satisfy differential privacy. Although this reduction in noise comes at the cost of some approximation error, we show that the benefits of the former significantly outweigh the latter. We explore the privacy-utility trade-off for several different in-stantiations of the proposed framework on two real-world data sets and show that useful social recommendations can be produced without sacrificing privacy. We also experimen-tally compare the proposed framework with several existing differential privacy mechanisms and show that the proposed framework significantly outperforms all of them in this set-ting. 1.

The idea of using one's behavior with a pointing device, such as a mouse or a touchpad, as a behavioral biometric for authentication purposes has gained increasing attention over the past decade. A number of interesting approaches based on the idea have emerged in the literature and promising experimental results have been reported; however, we argue that limitations in the past experimental evaluations of these approaches raise questions about their true effectiveness in a practical setting. In this paper, we review existing authentication approaches based on mouse dynamics and shed light on some important limitations regarding how the effectiveness of these approaches has been evaluated in the past. We present the results of several experiments that we conducted to illustrate our observations and suggest guidelines for evaluating future authentication approaches based on mouse dynamics. We also discuss a number of avenues for additional research that we believe are necessary to advance the state of the art in this area.

Statistical spam filters are known to be vulnerable to adversarial attacks. One such adversarial attack, known as the good word attack, thwarts spam filters by appending to spam messages sets of "good" words, which are common in legitimate e-mail but rare in spam. We present a counter attack strategy that first attempts to differentiate spam from legitimate e-mail in the input space, by transforming each e- mail into a bag of multiple segments, and subsequently applies multiple instance logistic regression on the bags. We treat each segment in the bag as an instance. An e-mail is classified as spam if at least one instance in the corresponding bag is spam, and as legitimate if all the instances in it are legitimate. We show that a spam filter using our multiple instance counter-attack strategy stands up better to good word attacks than its single instance counterpart and the commonly practiced Bayesian filters.

The minimum spanning tree clustering algorithm is known to be capable of detecting clusters with irregular boundaries. In this paper, we propose two minimum spanning tree based clustering algorithms. The first algorithm produces a k-partition of a set of points for any given k. The algorithm constructs a minimum spanning tree of the point set and removes edges that satisfy a predefined criterion. The process is repeated until k clusters are produced. The second algorithm partitions a point set into a group of clusters by maximizing the overall standard deviation reduction, without a given k value. We present our experimental results comparing our proposed algorithms to k-means and EM. We also apply our algorithms to image color clustering and compare our algorithms to the standard minimum spanning tree clustering algorithm