When analyzing confidential data through a privacy filter, a data scientist often needs to decide which queries will best support their intended analysis. For example, an analyst may wish to study noisy two-way marginals in a dataset produced by a mechanism M 1 . But, if the data are relatively sparse, the analyst may choose to examine noisy one-way marginals, produced by a mechanism M 2 , instead. Since the choice of whether to use M 1 or M 2 is data-dependent, a typical differentially private workflow is to first split the privacy loss budget ρ into two parts: ρ 1 and ρ 2 , then use the first part ρ 1 to determine which mechanism to use, and the remainder ρ 2 to obtain noisy answers from the chosen mechanism. In a sense, the first step seems wasteful because it takes away part of the privacy loss budget that could have been used to make the query answers more accurate. In this paper, we consider the question of whether the choice between M 1 and M 2 can be performed without wasting any privacy loss budget. For linear queries, we propose a method for decomposing M 1 and M 2 into three parts: (1) a mechanism M * that captures their shared information, (2) a mechanism M′1 that captures information that is specific to M 1 , (3) a mechanism M′2 that captures information that is specific to M 2 . Running M * and M′ 1 together is completely equivalent to running M 1 (both in terms of query answer accuracy and total privacy cost ρ ). Similarly, running M * and M′ 2 together is completely equivalent to running M 2 . Since M * will be used no matter what, the analyst can use its output to decide whether to subsequently run M ′ 1 (thus recreating the analysis supported by M 1 )or M′ 2 (recreating the analysis supported by M 2 ), without wasting privacy loss budget.

Recent advances in deep learning have led to superhuman performance across a variety of applications. Recently, these methods have been successfully employed to improve the rate-distortion performance in the task of image compression. However, current methods either use additional post-processing blocks on the decoder end to improve compression or propose an end-to-end compression scheme based on heuris-tics. For the majority of these, the trained deep neural networks (DNNs) are not compatible with standard encoders and would be difficult to deploy on personal com-puters and cellphones. In light of this, we propose a system that learns to improve the encoding performance by enhancing its internal neural representations on both the encoder and decoder ends, an approach we call Neural JPEG. We propose frequency domain pre-editing and post-editing methods to optimize the distribution of the DCT coefficients at both encoder and decoder ends in order to improve the stan-dard compression (JPEG) method. Moreover, we design and integrate a scheme for jointly learning quantization tables within this hybrid neural compression framework. In summary, our contributions are as follows:

Recent advances in deep learning have resulted in image compression algorithms that outperform JPEG and JPEG 2000 on the standard Kodak benchmark. However, they are slow to train (due to backprop-through-time) and, to the best of our knowledge, have not been systematically evaluated on a large variety of datasets. In this paper, we perform the first large scale comparison of recent state-of-the-art hybrid neural compression algorithms, while exploring the effects alternative training strategies (when applicable). The hybrid recurrent neural decoder is a former state-of-the-art model (recently overtaken by a Google model) that can be trained using backprop-through-time (BPTT) or with alternative algorithms like sparse attentive backtracking (SAB), unbiased online recurrent optimization (UORO), and real time recurrent learning (RTRL). We compare these training alternatives along with the Google models (GOOG and E2E) on 6 benchmark datasets. Surprisingly, we found that the model trained with SAB performs the better (outperforming even BPTT), resulting in faster convergence and better peak signal-to-noise ratio.

In practice, differentially private data releases are designed to support a variety of applications. A data release is fit for use if it meets target accuracy requirements for each application. In this paper, we consider the problem of answering linear queries under differential privacy subject to per-query accuracy constraints. Existing practical frameworks like the matrix mechanism do not provide such fine-grained control (they optimize total error, which allows some query answers to be more accurate than necessary, at the expense of other queries that become no longer useful). Thus, we design a fitness-for-use strategy that adds privacy-preserving Gaussian noise to query answers. The covariance structure of the noise is optimized to meet the fine-grained accuracy requirements while minimizing the cost to privacy.

The $L_1 $ regularization (Lasso) has proven to be a versatile tool to select relevant features and estimate the model coefficients simultaneously and has been widely used in many research areas such as genomes studies, finance, and biomedical imaging. Despite its popularity, it is very challenging to guarantee the feature selection consistency of Lasso especially when the dimension of the data is huge. One way to improve the feature selection consistency is to select an ideal tuning parameter. Traditional tuning criteria mainly focus on minimizing the estimated prediction error or maximizing the posterior model probability, such as cross-validation and BIC, which may either be time-consuming or fail to control the false discovery rate (FDR) when the number of features is extremely large. The other way is to introduce pseudo-features to learn the importance of the original ones. Recently, the Knockoff filter is proposed to control the FDR when performing feature selection. However, its performance is sensitive to the choice of the expected FDR threshold. Motivated by these ideas, we propose a new method using pseudo-features to obtain an ideal tuning parameter. In particular, we present the E fficient T uning of Lasso (ET-Lasso ) to separate active and inactive features by adding permuted features as pseudo-features in linear models. The pseudo-features are constructed to be inactive by nature, which can be used to obtain a cutoff to select the tuning parameter that separates active and inactive features. Experimental studies on both simulations and real-world data applications are provided to show that ET-Lasso can effectively and efficiently select active features under a wide range of scenarios.

Noisy Max and Sparse Vector are selection algorithms for differential privacy and serve as building blocks for more complex algorithms. In this paper we show that both algorithms can release additional information for free (i.e., at no additional privacy cost). Noisy Max is used to return the approximate maximizer among a set of queries. We show that it can also release for free the noisy gap between the approximate maximizer and runner-up. This free information can improve the accuracy of certain subsequent counting queries by up to 50%. Sparse Vector is used to return a set of queries that are approximately larger than a fixed threshold. We show that it can adaptively control its privacy budget (use less budget for queries that are likely to be much larger than the threshold) in order to increase the amount of queries it can process. These results follow from a careful privacy analysis.

The increased availability of large-scale trajectory data provides rich information for the study of urban dynamics. For example, New York City Taxi 8 Limousine Commission regularly releases source/destination information of taxi trips, where 173 million taxi trips released for Year 2013 [29]. Such a big dataset provides us potential new perspectives to address the traditional traffic problems. In this article, we study the travel time estimation problem. Instead of following the traditional route-based travel time estimation, we propose to simply use a large amount of taxi trips without using the intermediate trajectory points to estimate the travel time between source and destination. Our experiments show very promising results. The proposed big-data-driven approach significantly outperforms both state-of-the-art route-based method and online map services. Our study indicates that novel simple approaches could be empowered by big data and these approaches could serve as new baselines for some traditional computational problems.

Advances in sensor technology have enabled the collection of large-scale datasets. Such datasets can be extremely noisy and often contain a significant amount of outliers that result from sensor malfunction or human operation faults. In order to utilize such data for real-world applications, it is critical to detect outliers so that models built from these datasets will not be skewed by outliers. In this paper, we propose a new outlier detection method that utilizes the correlations in the data (e.g., taxi trip distance vs. trip time). Different from existing outlier detection methods, we build a robust regression model that explicitly models the outliers and detects outliers simultaneously with the model fitting. We validate our approach on real-world datasets against methods specifically designed for each dataset as well as the state of the art outlier detectors. Our outlier detection method achieves better performances, demonstrating the robustness and generality of our method. Last, we report interesting case studies on some outliers that result from atypical events.

Iterative algorithms, like gradient descent, are common tools for solving a variety of problems, such as model fitting. For this reason, there is interest in creating differentially private versions of them. However, their conversion to differentially private algorithms is often naive. For instance, a fixed number of iterations are chosen, the privacy budget is split evenly among them, and at each iteration, parameters are updated with a noisy gradient.

We consider the problem of privately releasing a class of queries that we call hierarchical count-of-counts histograms . Count-of-counts histograms partition the rows of an input table into groups (e.g., group of people in the same household), and for every integer j report the number of groups of size j . Hierarchical count-of-counts queries report count-of-counts histograms at different granularities as per hierarchy defined on an attribute in the input data (e.g., geographical location of a household at the national, state and county levels). In this paper, we introduce this problem, along with appropriate error metrics and propose a differentially private solution that generates count-of-counts histograms that are consistent across all levels of the hierarchy.

Page segmentation and table detection play an important role in understanding the structure of documents. We present a page segmentation algorithm that incorporates state-of-the-art deep learning methods for segmenting three types of document elements: text blocks, tables, and figures. We propose a multi-scale, multi-task fully convolutional neural network (FCN) for the tasks of semantic page segmentation and element contour detection. The semantic segmentation network accurately predicts the probability at each pixel of the three element classes. The contour detection network accurately predicts instance level "edges" around each element occurrence. We propose a conditional random field (CRF) that uses features output from the semantic segmentation and contour networks to improve upon the semantic segmentation network output. Given the semantic segmentation output, we also extract individual table instances from the page using some heuristic rules and a verification network to remove false positives. We show that although we only consider a page image as input, we produce comparable results with other methods that relies on PDF file information and heuristics and hand crafted features tailored to specific types of documents. Our approach learns the representative features for page segmentation from real and synthetic training data. %, and produces good results on real documents. The learning-based property makes it a more general method than existing methods in terms of document types and element appearances. For example, our method reliably detects sparsely lined tables which are hard for rule-based or heuristic methods.

We present an iterative refinement module that can be applied to the output feature maps of any existing convolutional neural networks in order to further improve classification accuracy. The proposed module, implemented by an attention-based recurrent neural network, can iteratively use its previous predictions to update attention and thereafter refine current predictions. In this way, the model is able to focus on a sub-region of input images to distinguish visually similar characters (see Figure 1 for an example). We evaluate its effectiveness on handwritten Chinese character recognition (HCCR) task and observe significant performance gain. HCCR task is challenging due to large number of classes and small differences between certain characters. To overcome these difficulties, we further propose a novel convolutional architecture that utilizes both low-level visual cues and high-level structural information. Together with the proposed iterative refinement module, our approach achieves an accuracy of 97.37%, outperforming previous methods that use raw images as input on ICDAR-2013 dataset [1].

The increased availability of large-scale trajectory data provides rich information for the study of urban dynamics. For example, New York City Taxi & Limousine Commission regularly releases source/destination information of taxi trips, where 173 million taxi trips released for Year 2013 [1]. Such a big dataset provides us potential new perspectives to address the traditional traffic problems. In this paper, we study the travel time estimation problem. Instead of following the traditional route-based travel time estimation, we propose to simply use a large amount of taxi trips without using the intermediate trajectory points to estimate the travel time between source and destination. Our experiments show very promising results. The proposed big data-driven approach significantly outperforms both state-of-the-art route-based method and online map services. Our study indicates that novel simple approaches could be empowered by the big data and these approaches could serve as new baselines for some traditional computational problems.

Crime is one of the most important social problems in the country, affecting public safety, children development, and adult socioeconomic status. Understanding what factors cause higher crime is critical for policy makers in their efforts to reduce crime and increase citizens' life quality. We tackle a fundamental problem in our paper: crime rate inference at the neighborhood level. Traditional approaches have used demographics and geographical influences to estimate crime rates in a region. With the fast development of positioning technology and prevalence of mobile devices, a large amount of modern urban data have been collected and such big data can provide new perspectives for understanding crime. In this paper, we used large-scale Point-Of-Interest data and taxi flow data in the city of Chicago, IL in the USA. We observed significantly improved performance in crime rate inference compared to using traditional features. Such an improvement is consistent over multiple years. We also show that these new features are significant in the feature importance analysis.

When analyzing data, it is important to account for all sources of noise. Public use datasets, such as those provided by the Census Bureau, often undergo additional perturbations designed to protect confidentiality. This source of noise is generally ignored in data analysis because crucial parameters and details about its implementation are withheld. In this paper, we consider the problem of inferring such parameters from the data. Specifically, we target data swapping, a perturbation technique commonly used by the U.S. Census Bureau and which, barring practical breakthroughs in disclosure control, will be used in the foreseeable future. The vanilla version of data swapping selects pairs of records and exchanges some of their attribute values. The number of swapped records is kept secret even though it is needed for data analysis and investigations into the confidentiality protection of individual records. We propose algorithms for estimating the number of swapped records in categorical data, even when the true data distribution is unknown.

When analyzing data that has been perturbed for privacy reasons, one is often concerned about its usefulness. Recent research on differential privacy has shown that the accuracy of many data queries can be improved by post-processing the perturbed data to ensure consistency constraints that are known to hold for the original data. Most prior work converted this post-processing step into a least squares minimization problem with customized efficient solutions. While improving accuracy, this approach ignored the noise distribution in the perturbed data.

In statistical privacy, utility refers to two concepts: information preservation, how much statistical information is retained by a sanitizing algorithm, and usability, how (and with how much difficulty) one extracts this information to build statistical models, answer queries, and so forth. Some scenarios incentivize a separation between information preservation and usability, so that the data owner first chooses a sanitizing algorithm to maximize a measure of information preservation, and, afterward, the data consumers process the sanitized output according to their various individual needs [Ghosh et al. 2009; Williams and McSherry 2010]. We analyze the information-preserving properties of utility measures with a combination of two new and three existing utility axioms and study how violations of an axiom can be fixed. We show that the average (over possible outputs of the sanitizer) error of Bayesian decision makers forms the unique class of utility measures that satisfy all of the axioms. The axioms are agnostic to Bayesian concepts such as subjective probabilities and hence strengthen support for Bayesian views in privacy research. In particular, this result connects information preservation to aspects of usability—if the information preservation of a sanitizing algorithm should be measured as the average error of a Bayesian decision maker, shouldn’t Bayesian decision theory be a good choice when it comes to using the sanitized outputs for various purposes? We put this idea to the test in the unattributed histogram problem where our decision-theoretic postprocessing algorithm empirically outperforms previously proposed approaches.

Data is a commodity. Recent research has considered the mathematical problem of setting prices for different queries over data. Ideal pricing functions need to be flexible --defined for arbitrary queries (select-project-join, aggregate, random sample, and noisy privacy-preserving queries). They should be fine-grained -- a consumer should not be required to buy the entire database to get answers to simple "low-information" queries (such as selecting only a few tuples or aggregating over only one attribute). Similarly, a consumer may not want to pay a large amount of money, only to discover that the database is empty. Finally, pricing functions should satisfy consistency conditions such as being "arbitrage-free" -- consumers should not be able to circumvent the pricing function by deducing the answer to an expensive query from a few cheap queries. Previously proposed pricing functions satisfy some of these criteria (i.e. they are defined for restricted subclasses of queries and/or use relaxed conditions for avoiding arbitrage). In this paper, we study arbitrage-free pricing functions defined for arbitrary queries. We propose new necessary conditions for avoiding arbitrage and provide new arbitrage-free pricing functions. We also prove several negative results related to the tension between flexible pricing and avoiding arbitrage, and show how this tension often results in unreasonable prices.

In this article, we introduce a new and general privacy framework called Pufferfish. The Pufferfish framework can be used to create new privacy definitions that are customized to the needs of a given application. The goal of Pufferfish is to allow experts in an application domain, who frequently do not have expertise in privacy, to develop rigorous privacy definitions for their data sharing needs. In addition to this, the Pufferfish framework can also be used to study existing privacy definitions. We illustrate the benefits with several applications of this privacy framework: we use it to analyze differential privacy and formalize a connection to attackers who believe that the data records are independent; we use it to create a privacy definition called hedging privacy, which can be used to rule out attackers whose prior beliefs are inconsistent with the data; we use the framework to define and study the notion of composition in a broader context than before; we show how to apply the framework to protect unbounded continuous attributes and aggregate information; and we show how to use the framework to rigorously account for prior data releases.

In statistical privacy, utility refers to two concepts: information preservation -- how much statistical information is retained by a sanitizing algorithm, and usability -- how (and with how much difficulty) does one extract this information to build statistical models, answer queries, etc. Some scenarios incentivize a separation between information preservation and usability, so that the data owner first chooses a sanitizing algorithm to maximize a measure of information preservation and, afterward, the data consumers process the sanitized output according to their needs [22, 46].

In this paper we introduce a new and general privacy framework called Pufferfish. The Pufferfish framework can be used to create new privacy definitions that are customized to the needs of a given application. The goal of Pufferfish is to allow experts in an application domain, who frequently do not have expertise in privacy, to develop rigorous privacy definitions for their data sharing needs. In addition to this, the Pufferfish framework can also be used to study existing privacy definitions.

Differential privacy is a powerful tool for providing privacy-preserving noisy query answers over statistical databases. It guarantees that the distribution of noisy query answers changes very little with the addition or deletion of any tuple. It is frequently accompanied by popularized claims that it provides privacy without any assumptions about the data and that it protects against attackers who know all but one record. In this paper we critically analyze the privacy protections offered by differential privacy.

Automatic recommendation of citations for a manuscript is highly valuable for scholarly activities since it can substantially improve the efficiency and quality of literature search. The prior techniques placed a considerable burden on users, who were required to provide a representative bibliography or to mark passages where citations are needed. In this paper we present a system that considerably reduces this burden: a user simply inputs a query manuscript (without a bibliography) and our system automatically finds locations where citations are needed. We show that naïve approaches do not work well due to massive noise in the document corpus. We produce a successful approach by carefully examining the relevance between segments in a query manuscript and the representative segments extracted from a document corpus. An extensive empirical evaluation using the CiteSeerX data set shows that our approach is effective.

This tutorial gives an overview of techniques for releasing data about individuals while preserving privacy.

"Privacy" and "utility" are words that frequently appear in the literature on statistical privacy. But what do these words really mean? In recent years, many problems with intuitive notions of privacy and utility have been uncovered. Thus more formal notions of privacy and utility, which are amenable to mathematical analysis, are needed. In this paper we present our initial work on an axiomatization of privacy and utility. In particular, we study how these concepts are affected by randomized algorithms. Our analysis yields new insights into the construction of both privacy definitions and mechanisms that generate data according to such definitions. In particular, it characterizes a class of relaxations of differential privacy and shows that desirable outputs of a differentially private mechanism are best interpreted as certain graphs rather than query answers or synthetic data.

When you write papers, how many times do you want to make some citations at a place but you are not sure which papers to cite? Do you wish to have a recommendation system which can recommend a small number of good candidates for every place that you want to make some citations? In this paper, we present our initiative of building a context-aware citation recommendation system. High quality citation recommendation is challenging: not only should the citations recommended be relevant to the paper under composition, but also should match the local contexts of the places citations are made. Moreover, it is far from trivial to model how the topic of the whole paper and the contexts of the citation places should affect the selection and ranking of citations. To tackle the problem, we develop a context-aware approach. The core idea is to design a novel non-parametric probabilistic model which can measure the context-based relevance between a citation context and a document. Our approach can recommend citations for a context effectively. Moreover, it can recommend a set of citations for a paper with high quality. We implement a prototype system in CiteSeerX. An extensive empirical evaluation in the CiteSeerX digital library against many baselines demonstrates the effectiveness and the scalability of our approach.

In this paper we present a method for reasoning about privacy using the concepts of exchangeability and deFinetti's theorem. We illustrate the usefulness of this technique by using it to attack a popular data sanitization scheme known as Anatomy. We stress that Anatomy is not the only sanitization scheme that is vulnerable to this attack. In fact, any scheme that uses the random worlds model, i.i.d. model, or tuple-independent model needs to be re-evaluated.

In this paper, we propose the first formal privacy analysis of a data anonymization process known as the synthetic data generation, a technique becoming popular in the statistics community. The target application for this work is a mapping program that shows the commuting patterns of the population of the United States. The source data for this application were collected by the U.S. Census Bureau, but due to privacy constraints, they cannot be used directly by the mapping program. Instead, we generate synthetic data that statistically mimic the original data while providing privacy guarantees. We use these synthetic data as a surrogate for the original data. We find that while some existing definitions of privacy are inapplicable to our target application, others are too conservative and render the synthetic data useless since they guard against privacy breaches that are very unlikely. Moreover, the data in our target application is sparse, and none of the existing solutions are tailored to anonymize sparse data. In this paper, we propose solutions to address the above issues.

We study how best to schedule scans of large data files, in the presence of many simultaneous requests to a common set of files. The objective is to maximize the overall rate of processing these files, by sharing scans of the same file as aggressively as possible, without imposing undue wait time on individual jobs. This scheduling problem arises in batch data processing environments such as Map-Reduce systems, some of which handle tens of thousands of processing requests daily, over a shared set of files. As we demonstrate, conventional scheduling techniques such as shortest-job-first do not perform well in the presence of cross-job sharing opportunities. We derive a new family of scheduling policies specifically targeted to sharable workloads. Our scheduling policies revolve around the notion that, all else being equal, it is good to schedule nonsharable scans ahead of ones that can share IO work with future jobs, if the arrival rate of sharable future jobs is expected to be high. We evaluate our policies via simulation over varied synthetic and real workloads, and demonstrate significant performance gains compared with conventional scheduling approaches.

Recent work has shown the necessity of considering an attacker's background knowledge when reasoning about privacy in data publishing. However, in practice, the data publisher does not know what background knowledge the attacker possesses. Thus, it is important to consider the worst-case. In this paper, we initiate a formal study of worst-case background knowledge. We propose a language that can express any background knowledge about the data. We provide a polynomial time algorithm to measure the amount of disclosure of sensitive information in the worst case, given that the attacker has at most k pieces of information in this language. We also provide a method to efficiently sanitize the data so that the amount of disclosure in the worst case is less than a specified threshold.

Publishing data about individuals without revealing sensitive information about them is an important problem. In recent years, a new definition of privacy called k -anonymity has gained popularity. In a k -anonymized dataset, each record is indistinguishable from at least k − 1 other records with respect to certain identifying attributes. In this article, we show using two simple attacks that a k -anonymized dataset has some subtle but severe privacy problems. First, an attacker can discover the values of sensitive attributes when there is little diversity in those sensitive attributes. This is a known problem. Second, attackers often have background knowledge, and we show that k -anonymity does not guarantee privacy against attackers using background knowledge. We give a detailed analysis of these two attacks, and we propose a novel and powerful privacy criterion called ℓ-diversity that can defend against such attacks. In addition to building a formal foundation for ℓ-diversity, we show in an experimental evaluation that ℓ-diversity is practical and can be implemented efficiently.

Publishing data about individuals without revealing sensitive information about them is an important problem. In recent years, a new definition of privacy called \kappa-anonymity has gained popularity. In a \kappa-anonymized dataset, each record is indistinguishable from at least k—1 other records with respect to certain "identifying" attributes. In this paper we show with two simple attacks that a \kappa-anonymized dataset has some subtle, but severe privacy problems. First, we show that an attacker can discover the values of sensitive attributes when there is little diversity in those sensitive attributes. Second, attackers often have background knowledge, and we show that \kappa-anonymity does not guarantee privacy against attackers using background knowledge. We give a detailed analysis of these two attacks and we propose a novel and powerful privacy definition called \ell-diversity. In addition to building a formal foundation for \ell-diversity, we show in an experimental evaluation that \ell-diversity is practical and can be implemented efficiently.

Limiting disclosure in data publishing requires a careful balance between privacy and utility. Information about individuals must not be revealed, but a dataset should still be useful for studying the characteristics of a population. Privacy requirements such as k-anonymity and l-diversity are designed to thwart attacks that attempt to identify individuals in the data and to discover their sensitive information. On the other hand, the utility of such data has not been well-studied.In this paper we will discuss the shortcomings of current heuristic approaches to measuring utility and we will introduce a formal approach to measuring utility. Armed with this utility metric, we will show how to inject additional information into k-anonymous and l-diverse tables. This information has an intuitive semantic meaning, it increases the utility beyond what is possible in the original k-anonymity and l-diversity frameworks, and it maintains the privacy guarantees of k-anonymity and l-diversity.

The subfield of itemset mining is essentially a collection of algorithms. Whenever a new type of constraint is discovered, a specialized algorithm is proposed to handle it. All of these algorithms are highly tuned to take advantage of the unique properties of their associated constraints, and so they are not very compatible with other constraints. In this paper we present a more unified view of mining constrained itemsets such that most existing algorithms can be easily extended to handle constraints for which they were not designed a-priori. We apply this technique to mining itemsets with restrictions on their variance --- a problem that has been open for several years in the data mining community.

Constraint-based mining of itemsets for questions such as "find all frequent itemsets where the total price is at least $50" has received much attention recently. Two classes of constraints, monotone and antimonotone, have been identified as very useful. There are algorithms that efficiently take advantage of either one of these two classes, but no previous algorithms can efficiently handle both types of constraints simultaneously. In this paper, we present the first algorithm (called DualMiner) that uses both monotone and antimonotone constraints to prune its search space. We complement a theoretical analysis and proof of correctness of DualMiner with an experimental study that shows the efficacy of DualMiner compared to previous work.