Power consumption anomaly detection systems that use neural networks for prediction tasks are vulnerable to adversarial attacks, leading to unreliable performance and potential adverse effects on the power grid. Certain attack configurations can evade detection or trigger false alarms due to the neural networks’ generalization tendencies, causing adaptation to attack values. This study investigates the effectiveness of adversarial training methods in improving detection performance against such attacks on power consumption data. Leveraging techniques like the Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and Projected Gradient Descent (PGD), we assess the resilience of anomaly detection models. Through empirical experiments, we evaluate detection accuracy, adjustment capabilities, and prediction errors of these adversarially trained models across three datasets. Our results show significant improvements in detection performance, particularly in attack scenarios that normal prediction models would not detect. Additionally, we analyze the models’ adaptability to anomalous data and quantify prediction errors, providing insights into their robustness and limitations. Integrating adversarial training techniques into anomaly detection models for power grids can reduce over-generalization to attack data, enhancing the detection of malicious demand manipulation attacks.

Privacy policies outline data collection and sharing practices followed by an organization, together with choice and control measures available to users to manage the process. However, users have often needed help reading and understanding such documents, regardless of their being written in a natural language. The fundamental problems with privacy policies persist despite advancements in privacy design, frameworks, and regulations. To identify the causes of privacy policies being persistently challenging to comprehend, it is vital to investigate historical policy patterns and understand the evolution of privacy policies concerning information packaging and presentation. To this aid, we create a sentence-level classifier to conduct a large-scale longitudinal analysis on different privacy policies from 130,604 organizations, totaling approximately one million policies from 1997 to 2019. We annotate 10,717 sentences from 115 policies in the OPP-115 corpus to implement the classifier and then use those annotations to train the XLNet and BERT classifiers. Results from our analysis reveal that specific data practice categories experience more frequent policy changes than others, making it challenging to track relevant information over time. In addition, we discover that every category has distinct composition, readability, and structural issues, which exacerbate when categories frequently co-occur in a document. Based on our observations, we provide recommendations for policy articulation and revision to make privacy policy documents conform to better coherence and structure.

Privacy policies inform users of the data practices and access protocols employed by organizations and their digital counterparts. Research has shown that users often feel that these privacy policies are lengthy and complex to read and comprehend. However, it is critical for people to be aware of the data access practices employed by the organizations. Hence, much research has focused on automatically extracting privacy-specific artifacts from the policies, predominantly by using natural language classification tools. However, these classification tools are designed primarily for the classification of paragraphs or segments of the policies. In this paper, we report on our research where we identify the gap in classifying policies at a segment level, and provide an alternate definition of segment classification using sentence classification. To this aid, we train and evaluate sentence classifiers for privacy policies using BERT and XLNet. Our approach demonstrates improvements in prediction quality of existing models and hence, surpasses the current baselines for classification models, without requiring additional parameter and model tuning. Using our sentence classifiers, we also study topical structures in Alexa top 5000 website policies, in order to identify and quantify the diffusion of information pertaining to privacy-specific topics in a policy.

Privacy policies notify users about the privacy practices of websites, mobile apps, and other products and services. However, users rarely read them and struggle to understand their contents. Due to the complicated nature of these documents, it gets even harder to understand and take note of any changes of interest or concern when the policies are changed or revised. With advances in machine learning and natural language processing, tools that can automatically annotate sentences of policies have been developed. These annotations can help a user identify and understand relevant parts of a privacy policy. In this paper, we present our attempt to further such annotations by also detecting the important changes that occurred across sentences. Using supervised machine learning models, word-embedding, similarity matching, and structural analysis of sentences, we present a process that takes two different versions of a privacy policy as input, matches the sentences of one version to another based on semantic similarity, and identifies relevant changes between two matched sentences. We present the results and insights of applying our approach on 79 privacy policies manually downloaded from Facebook, WhatsApp, Twitter, Google, LinkedIn and Snapchat, ranging between the period of 1999 to 2020.

An increased usage in IoT devices across the globe has posed a threat to the power grid. When an attacker has access to multiple IoT devices within the same geographical location, they can possibly disrupt the power grid by regulating a botnet of high-wattage IoT devices. Anomaly detection comes handy to inform the power operator of an anomalous behavior during such an attack. However, it is difficult to detect anomalies when attacks take place obscurely and for prolonged time periods. To effectively detect such attacks, we propose a novel dynamic thresholding mechanism that is used with prediction-based anomaly score techniques. We compare our detection rates to predefined thresholding mechanisms and commercial detection methods and observe that our method improves the detection rate up to 97% across different attacks that we generate.

Abstract The mobile eco-system is driven by an increasing number of location-aware applications. Consequently, a number of location privacy models have been proposed to prevent the unwanted inference of sensitive information from location traces. A primary focus in these models is to ensure that a privacy mechanism can indeed retrieve results that are geographically the closest. However, geo-query results are, in most cases, ranked using a combination of distance and importance data, thereby producing a result landscape that is periodically flat and not always dictated by distance. A privacy model that does not exploit this structure of geo-query results may enforce weaker levels of location privacy. Towards this end, we explore a formal location privacy principle designed to capture arbitrary similarity between locations, be it distance, or the number of objects common in their result sets. We propose a composite privacy mechanism that performs probabilistic cloaking and exponentially weighted sampling to provide coarse grain location hiding within a tunable area, and finer privacy guarantees under the principle inside this area. We present extensive empirical evidence to supplement claims on the effectiveness of the approach, along with comparative results to assert the stronger privacy guarantees.

Location privacy preservation algorithms in the context of location-based services have evolved in the recent years. However, a majority of the proposals assume that points of interests (POI) are ranked only by distance, and demand extensive architectural changes. As a result, a significant gap remains between academic proposals and the industry standard of implementing location based services. Recent advances in mobile device capabilities, more specifically in their computational power and energy efficiency, have opened the possibility of engaging the client hardware more actively in the execution of a privacy algorithm, thereby relaxing strong dependencies on trusted third parties or the service provider. With this motivation, we propose a novel privacy algorithm that determines the most prominent result set through operations restricted to the client device, thereby limiting the communication of precise location information to the service provider. The service provider only acts as a data source, and is required to perform operations that are within existing industry norms. By measuring the privacy offered by the algorithm under a formal threat model, we demonstrate its robustness and practicability, and supplement our conclusions with empirical evidence.

Abstract A federated query portal in an electronic health record infrastructure enables large epidemiology studies by combining data from geographically dispersed medical institutions. However, an individual’s health record has been found to be distributed across multiple carrier databases in local settings. Privacy regulations may prohibit a data source from revealing clear text identifiers, thereby making it non-trivial for a query aggregator to determine which records correspond to the same underlying individual. In this paper, we explore this problem of privately detecting and tracking the health records of an individual in a distributed infrastructure. We begin with a secure set intersection protocol based on commutative encryption, and show how to make it practical on comparison spaces as large as 1010 pairs. Using bigram matching, precomputed tables, and data parallelism, we successfully reduced the execution time to a matter of minutes, while retaining a high degree of accuracy even in records with data entry errors. We also propose techniques to prevent the inference of identifier information when knowledge of underlying data distributions is known to an adversary. Finally, we discuss how records can be tracked utilizing the detection results during query processing.

Location privacy preservation algorithms for nearby points-of-interest (POI) search have evolved in the recent years. However, a majority of the proposals assume that points of interests are ranked only by distance, and demand extensive architectural changes. As a result, a significant gap remains between academic proposals and the industry standard of implementing location based services. Recent advances in mobile device capabilities, more specifically in their computational power and energy efficiency, have opened the possibility of engaging the client hardware more actively in the execution of a privacy algorithm, thereby relaxing strong dependencies on trusted third parties or the service provider. With this motivation, we propose a novel privacy algorithm for use in POI search that achieves much of the desired location privacy by restricting the usage of precise location data to the client device.

When aggregating medical data for research, it is necessary to link data on the same person, but from different sources. Linking enables a researcher to conduct longitudinal studies. Typically such linking can be accomplished by using personal identifying information, such as names, birthdates, addresses, and national or local identifying codes, though occasionally this method does not work because of incompleteness or inaccuracies in the data. For research, the Health Insurance Portability and Accountability Act (HIPAA) privacy rules severely restrict researcher access to identifiers. Therefore, an important research problem is how to link data from a geographic region whose data sources have significant overlap in the actual patients included. In this talk, I describe various challenges and opportunities that exist while tackling this problem.

Location-based applications utilize the positioning capabilities of a mobile device to determine the current location of a user, and customize query results to include neighboring points of interests. However, location knowledge is often perceived as personal information. One of the immediate issues hindering the wide acceptance of location-based applications is the lack of appropriate methodologies that offer fine grain privacy controls to a user without vastly affecting the usability of the service. While a number of privacy-preserving models and algorithms have taken shape in the past few years, there is an almost universal need to specify one's privacy requirement without understanding its implications on the service quality. In this paper, we propose a user-centric location-based service architecture where a user can observe the impact of location inaccuracy on the service accuracy before deciding the geo-coordinates to use in a query. We construct a local search application based on this architecture and demonstrate how meaningful information can be exchanged between the user and the service provider to allow the inference of contours depicting the change in query results across a geographic area. Results indicate the possibility of large default privacy regions (areas of no change in result set) in such applications.

Location privacy research has received wide attention in the past few years owing to the growing popularity of location-based applications, and the skepticism thereof on the collection of location information. A large section of this research is directed toward mechanisms based on location obfuscation enforced using cloaking regions. The primary motivation for this engagement comes from the relatively well-researched area of database privacy. Researchers in this sibling domain have indicated multiple times that any notion of privacy is incomplete without explicit statements on the capabilities of an adversary. As a result, we have started to see some efforts to categorize the various forms of background knowledge that an adversary may possess in the context of location privacy. Along this line, we consider some preliminary forms of attacker knowledge, and explore what implication does a certain form of knowledge has on location privacy. Continuing on, we extend our insights to a form of adversarial knowledge related to the geographic uncertainty that the adversary has in correctly locating a user. We empirically demonstrate that the use of cloaking regions can adversely impact the preservation of privacy in the presence of such approximate location knowledge, and demonstrate how perturbation-based mechanisms can instead provide a well-balanced tradeoff between privacy and service accuracy.

Security risk assessment and mitigation are two vital processes that need to be executed to maintain a productive IT infrastructure. On one hand, models such as attack graphs and attack trees have been proposed to assess the cause-consequence relationships between various network states, while on the other hand, different decision problems have been explored to identify the minimum-cost hardening measures. However, these risk models do not help reason about the causal dependencies between network states. Further, the optimization formulations ignore the issue of resource availability while analyzing a risk model. In this paper, we propose a risk management framework using Bayesian networks that enable a system administrator to quantify the chances of network compromise at various levels. We show how to use this information to develop a security mitigation and management plan. In contrast to other similar models, this risk model lends itself to dynamic analysis during the deployed phase of the network. A multiobjective optimization platform provides the administrator with all trade-off information required to make decisions in a resource constrained environment.

Data anonymization techniques have received extensive attention in the privacy research community over the past several years. Various models of privacy preservation have been proposed: k-anonymity, ℓ-diversity and t-closeness, to name a few. An oft-cited drawback of these models is that there is considerable loss in data quality arising from the use of generalization and suppression techniques. Optimization attempts in this context have so far focused on maximizing the data utility for a pre-specified level of privacy. To determine if better privacy levels are obtainable with the same level of data utility, majority of the existing formulations require exhaustive analysis. Further, the data publisher's perspective is often missed in the process. The publisher wishes to maintain a given level of data utility (since the data utility is the revenue earner) and then maximize the level of privacy within acceptable limits. In this paper, we explore this privacy versus data quality trade-off as a multi-objective optimization problem. Our goal is to provide substantial information to a data publisher about the trade-offs available between the privacy level and the information content of an anonymized data set.

Privacy constraints are typically enforced on shared data that contain sensitive personal attributes. However, owing to its adverse effect on the utility of the data, information loss must be minimized while sanitizing the data. Existing methods for this purpose modify the data only to the extent necessary to satisfy the privacy constraints, thereby asserting that the information loss has been minimized. However, given the subjective nature of information loss, it is often difficult to justify such an assertion. In this paper, we propose an interactive procedure to generate a data generalization scheme that optimally meets the preferences of the data publisher. A data publisher guides the sanitization process by specifying aspirations in terms of desired achievement levels in the objectives. A reference direction based methodology is used to investigate neighborhood solutions if the generated scheme is not acceptable. This approach draws its power from the constructive input received from the publisher about the suitability of a solution before finding a new one.

Location obfuscation using cloaking regions preserves location anonymity by hiding the true user among a set of other equally likely users. Furthermore, a cloaking region should also guarantee that the type of queries issued by users within the region are mutually diverse enough. The first requirement is fulfilled by satisfying location k-anonymity while the second one is ensured by satisfying query l-diversity. However, these two models are not sufficient to prevent the association of queries to users when the service depends on continuous location updates. Successive cloaking regions for a user may be k-anonymous and query l-diverse but still be prone to correlation attacks. In this paper, we provide a formal analysis of the privacy risks involved in a continuous location-based service, and show how continuous queries can invalidate the privacy guarantees provided by k-anonymity and l-diversity. Drawing upon the principle of m-invariance in database privacy, we show how query m-invariance can provide location and query privacy in continuous services.

Data generalization is widely used to protect identities and prevent inference of sensitive information during the public release of microdata. The k-anonymity model has been extensively applied in this context. The model seeks a generalization scheme such that every individual becomes indistinguishable from at least k-1 other individuals and the loss in information while doing so is kept at a minimum. The search is performed on a domain hierarchy lattice where every node is a vector signifying the level of generalization for each attribute. An effort to understand privacy and data utility trade-offs will require knowing the minimum possible information losses of every possible value of k. However, this can easily lead to an exhaustive evaluation of all nodes in the hierarchy lattice. In this paper, we propose using the concept of Pareto-optimality to obtain the desired trade-off information. A Pareto-optimal generalization is one in which no other generalization can provide a higher value of k without increasing the information loss. We introduce the Pareto-Optimal k-Anonymization (POkA) algorithm to traverse the hierarchy lattice and show that the number of node evaluations required to find the Pareto-optimal generalizations can be significantly reduced. Results on a benchmark data set show that the algorithm is capable of identifying all Pareto-optimal nodes by evaluating only 20% of nodes in the lattice.

Privacy models such as k-anonymity and l-diversity typically offer an aggregate or scalar notion of the privacy property that holds collectively on the entire anonymized data set. However, they fail to give an accurate measure of privacy with respect to the individual tuples. For example, two anonymizations achieving the same value of k in the k-anonymity model will be considered equally good with respect to privacy protection. However, it is quite possible that for one of the anonymizations a majority of the individual tuples have lesser probabilities of privacy breaches than their counterparts in the other anonymization. We therefore reject the notion that all anonymizations satisfying a particular privacy property, such as k-anonymity, are equally good. The scalar or aggregate value used in privacy models is often biased towards a fraction of the data set, resulting in higher privacy for some individuals and minimalistic for others. Consequently, to better compare anonymization algorithms, there is a need to formalize and measure this bias. Towards this end, we advocate the use of vector-based methods for representing privacy and other measurable properties of an anonymization. We represent the measure of a given property for an anonymized data set using a property vector. Anonymizations are then compared using quality index functions that quantify the effectiveness of the property vectors. A formal analysis with respect to their scope and limitations is provided. Finally, we present preference based techniques when comparisons are to be made across multiple properties induced by anonymizations.

Public data sharing is utilized in a number of businesses to facilitate the exchange of information. Privacy constraints are usually enforced to prevent unwanted inference of information, specially when the shared data contain sensitive personal attributes. This, however, has an adverse effect on the utility of the data for statistical studies. Thus, a requirement while modifying the data is to minimize the information loss. Existing methods employ the notion of "minimal distortion" where the data is modified only to the extent necessary to satisfy the privacy constraint, thereby asserting that the information loss has been minimized. However, given the subjective nature of information loss, it is often difficult to justify this assertion. In this paper, we propose an evolutionary algorithm to explicitly minimize an achievement function given constraints on the privacy level of the transformed data. Privacy constraints specified in terms of anonymity models are modeled as additional objectives and an evolutionary multi-objective approach is proposed. We highlight the requirement to minimize any bias induced by the anonymity model and present a scalarization incorporating preferences in information loss and privacy bias as the achievement function.

Data dissemination in pervasive environments is often accomplished by on-demand broadcasting. The time critical nature of the data requests plays an important role in scheduling these broadcasts. Most research in on-demand broadcast scheduling has focused on the timely servicing of requests so as to minimize the number of missed deadlines. However, there exists many pervasive environments where the utility of the data is an equally important criterion as its timeliness. Missing the deadline reduces the utility of the data but does not make it zero. In this work, we address the problem of scheduling on-demand data broadcasts with soft deadlines. We investigate search based optimization techniques to develop broadcast schedulers that make explicit attempts to maximize the utility of data requests as well as service as many requests as possible within the acceptable time limit. Our analysis shows that heuristic driven methods for such problems can be improved by hybridizing them with local search algorithms. We further investigate the option of employing a dynamic optimization technique to facilitate utility gain, thereby surpassing the requirement of a heuristic in the process. An evolution strategy based stochastic hill climber is investigated in this context.

Data broadcasting makes effective use of low bandwidth and is commonly used in applications involving mobile devices. We consider the case where data must be broadcast in a particular order and within a specified response time. However, communication bottlenecks prohibit the timely serving of all requests; although, missing the deadline does not make the data utility zero. In this work, we consider the problem of real-time data broadcast scheduling in the presence of soft deadlines together with constraints on the order in which data-items should be broadcast to be useful. We explore the method of evolution strategy to solve the problem, keeping in view that the real-time scheduler has to effectively trade-off between its running time and the quality of schedules generated.

When disseminating data involving human subjects, researchers have to weigh in the requirements of privacy of the individuals involved in the data. A model widely used for enhancing individual privacy is k-anonymity, where an individual data record is rendered similar to k - 1 other records in the data set by using generalization and/or suppression operations on the data attributes. The drawback of this model is that such transformations result in considerable loss of information that is proportional to the choice of k. Studies in this context have so far focused on minimizing the information loss for some given value of k. However, owing to the presence of outliers, a specified k value may or may not be obtainable. Further, an exhaustive analysis is required to determine a k value that fits the loss constraint specified by a data publisher. In this paper, we formulate a multi-objective optimization problem to illustrate that the decision on k can be much more informed than being a choice solely based on the privacy requirement. The optimization problem is intended to resolve the issue of data privacy when data suppression is not allowed in order to obtain a particular value of k. An evolutionary algorithm is employed here to provide this insight.

Researchers have previously looked into the problem of determining if a given set of security hardening measures can effectively make a networked system secure. Many of them also addressed the problem of minimizing the total cost of implementing these hardening measures, given costs for individual measures. However, system administrators are often faced with a more challenging problem since they have to work within a fixed budget which may be less than the minimum cost of system hardening. Their problem is how to select a subset of security hardening measures so as to be within the budget and yet minimize the residual damage to the system caused by not plugging all required security holes. In this work, we develop a systematic approach to solve this problem by formulating it as a multi-objective optimization problem on an attack tree model of the system and then use an evolutionary algorithm to solve it.

Stencil computations form the performance-critical core of many applications. Tiling and parallelization are two important optimizations to speed up stencil computations. Many tiling and parallelization strategies are applicable to a given stencil computation. The best strategy depends not only on the combination of the two techniques, but also on many parameters: tile and loop sizes in each dimension; computation-communication balance of the code; processor architecture; message startup costs; etc. The best choices can only be determined through design-space exploration, which is extremely tedious and error prone to do via exhaustive experimentation. We characterize the space of multi-level tilings and parallelizations for 2D/3D Gauss-Siedel stencil computation. A systematic exploration of a part of this space enabled us to derive a design which is up to a factor of two faster than the standard implementation.

Heterogeneous distributed computing systems often must operate in an environment where system parameters are subject to uncertainty. Robustness can be defined as the degree to which a system can function correctly in the presence of parameter values different from those assumed. We present a methodology for quantifying the robustness of resource allocations in a dynamic environment where task execution times are stochastic. The methodology is evaluated through measuring the robustness of three different resource allocation heuristics within the context of a stochastic dynamic environment. A Bayesian regression model is fit to the combined results of the three heuristics to demonstrate the correlation between the stochastic robustness metric and the presented performance metric. The correlation results demonstrated the significant potential of the stochastic robustness metric to predict the relative performance of the three heuristics given a common objective function.