As Android Automotive OS (AAOS) becomes the in-vehicle platform of choice for infotainment and domain-controller functions in modern passenger cars, its software supply chain has emerged as a critical security frontier. AAOS spans both infotainment and vehicle-control domains within the automotive electronics architecture by supporting media streaming, over-the-air updates, navigation, and sensor fusion. Its open-source foundations and reliance on third-party libraries introduce risks, from outdated components to malicious modules, that can undermine vehicle functionality and passenger safety. In recognition of these threats, ISO/SAE 21434 and UNECE WP.29 R155 mandate structured security assessments for vehicular systems to prevent software-chain vulnerabilities from compromising safety. In this study, we apply a shift-right security analysis via firmware reverse engineering to AAOS images from four leading OEMs. We unpack each firmware image, extract software bills of materials (SBOMs), map Common Vulnerabilities and Exposures (CVE) to components, and characterize system-level attack surfaces across infotainment and control subsystems. Proof-of-concept exploits were developed for high-risk vulnerabilities. One critical CVE was successfully triggered, while others were mitigated by missing dependencies or built-in protections. Our work delivers a reproducible firmware-analysis workflow for automotive supply-chain risk assessment, a comparative survey of third-party and proprietary component management, and the evidence of inconsistent security postures in AAOS-based vehicular electronics. These vulnerabilities underscore the need for harmonized SBOM practices and targeted hardening in next-generation in-vehicle systems.

• Introduces PPO-CIS, a novel Deep Reinforcement Learning (DRL) framework for adaptive toxicity detection in classifier cascades. • Utilizes Proximal Policy Optimization (PPO) to dynamically select classifiers based on sample complexity and system cost constraints. • Proposes a custom reward function balancing classification accuracy, latency, and computational efficiency. • Demonstrates improved performance over static cascades and individual models across two benchmark datasets: Kaggle and ToxiGen. • Achieves significant gains in throughput and accuracy, making the system suitable for real-time content moderation at scale. • Provides a scalable and cost-effective moderation solution for social media platforms facing high data volume and regulatory pressure. Online platforms face growing challenges in moderating harmful user-generated content due to the large volume and rapid pace of interactions. In existing moderation systems, automated tools assist human moderators, yet they often struggle to balance processing efficiency and reliable classification. When moderation fails to detect harmful content quickly and accurately, platforms risk user harm and noncompliance with content safety regulations. This paper proposes an adaptive moderation method named the Proximal Policy Optimization-based Cascaded Inference System (PPO-CIS). The method integrates multiple toxicity classifiers into a cascaded decision architecture guided by deep reinforcement learning. At each step, PPO-CIS selects the next classifier according to the content difficulty and the expected gain in accuracy relative to computational cost. The system enables rapid filtering of benign content and only activates high-capacity models for uncertain cases. PPO-CIS is the first toxicity detection framework to employ Proximal Policy Optimization for real-time optimization of classifier cascades. Experiments on the AugmenToxic and ToxiGen datasets show that PPO-CIS improves detection accuracy by 2.10 percent while increasing processing speed from 42.74 to 384 samples per second compared with static cascade designs. The findings show that adaptive model selection can better shield users from exposure to harmful content while lowering moderation costs. PPO-CIS provides a practical solution for deploying scalable and timely content moderation in fast-moving online environments.

Researchers often approach malware detection as a binary classification problem. However, evidence indicates that malware can belong to multiple families simultaneously, and malicious files frequently exhibit numerous benign features. Attackers exploit this by embedding malicious intent within benign features, making malware detection a problem better suited for fuzzy systems. Furthermore, providing explainability for such fuzzy classification remains a significant challenge, requiring specialized Explainable AI (XAI) frameworks. Existing XAI approaches offer insights into model decisions but are vulnerable to adversarial attacks that manipulate features to mislead models. To address these issues, we propose PAC-X, a novel XAI framework for malware detection. PAC-X integrates the Conditional Attention Neural Network (CAN-Net) to deliver comprehensive multi-fuzzy-class explainability and employs Contextual Fuzzy Clustering (CFC) to extract contextual insights from training data. This framework is resilient to adversarial manipulations, maintaining reliable and interpretable explanations even under adversarial conditions designed to mislead the model. Through extensive evaluations on diverse malware datasets, PAC-X demonstrates superior explainability and robustness compared to state-of-the-art XAI methods. It provides a critical advancement in cybersecurity by addressing the complexities of evasive malware detection and enabling a deeper interpretation of multi-class malware characteristics.

Cookie synchronization enables multiple advertising networks to share user browsing data, thus refining ad targeting without explicit user consent. Although existing mitigation strategies, such as cookie blockers, reduce privacy risks, they can also decrease website revenue or restrict certain online services. To address this limitation, we introduce a differentially private cookie synchronization proxy, CSyncProxy, that leverages the exponential mechanism to grant users granular control over the amount of data shared during cookie synchronization. By obfuscating user identities across websites, the proxy maintains sufficient personalization for advertisements while safeguarding user privacy. Experimental web crawls of the top 100 websites per visit and analysis of the likelihood of successful anonymization indicate that our approach reduces the number of instances of cookie synchronization by over 40% compared to standard browsing and 11% compared to Chrome’s cookie-blocking feature, without blocking or barring the user from any website.

User identity linkage (UIL) is the task of aligning user identities of the same user across different social network platforms. Although existing approaches have explored various aspects such as different user profile attributes and social network structures, the writing styles from user-generated texts, which is commonly known as stylometry, remain relatively underexplored. In this paper, we propose a novel Graph Neural Network (GNN)-based model named StyleLink, which leverages both social network structures and stylometric features derived from user-generated texts to address the UIL problem in an integrated manner. Our model utilizes GNNs to incorporate both stylometric features and the network structure for each social network, effectively embedding the network and enhancing user representation. This is the first work to incorporate stylometric features into GNNs to embed social networks and then conduct UIL between two embedding spaces. Extensive experiments on real-world social network datasets demonstrate the superior performance of StyleLink over existing state-of-the-art methods, achieving higher accuracy in user linkage and improved ranking of identity matches. In addition, we explore the effects of different linguistic characteristics in the identification of user identities and visualizes the effects of applying GNNs for better social network embedding.

Pretraining large language models is a complex endeavor influenced by multiple factors, including model architecture, data quality, training continuity, and hardware constraints. In this paper, we share insights gained from the experience of training DMaS-LLaMa-Lite, a fully open source, 1.7-billionparameter, LLaMa-based model, on approximately 20 billion tokens of carefully curated data. We chronicle the full training trajectory, documenting how evolving validation loss levels and downstream benchmarks reflect transitions from incoherent text to fluent, contextually grounded output. Beyond pretraining, we extend our analysis to include a post-training phase focused on instruction tuning, where the model was refined to produce more contextually appropriate, user-aligned responses. We highlight practical considerations such as the importance of restoring optimizer states when resuming from checkpoints, and the impact of hardware changes on training stability and throughput. While qualitative evaluation provides an intuitive understanding of model improvements, our analysis extends to various performance benchmarks, demonstrating how high-quality data and thoughtful scaling enable competitive results with significantly fewer training tokens. By detailing these experiences and offering training logs, checkpoints, and sample outputs, we aim to guide future researchers and practitioners in refining their pretraining strategies.

This special day event on Responsible Artificial Intelligence (AI) brings together researchers, practitioners, and policymakers to explore how data mining and machine learning systems can be designed to align with ethical principles, societal values, and human well-being. As AI technologies increasingly influence decisions in healthcare, finance, governance, and social systems, there is a critical need to develop frameworks that embed fairness, accountability, and privacy directly into the foundations of knowledge discovery. This full-day event will feature a mix of invited talks, interactive debates, expert panels, and peer-reviewed research presentations, all focused on the practical integration of ethical design into data-driven systems. The Responsible AI Day builds on the success of Canada's NSERC CREATE Program on Responsible AI, an interdisciplinary initiative training the next generation of AI researchers across computer science, law, bioethics, public health, and media studies. Topics will span scalable AI governance, privacy-preserving computation, algorithmic bias mitigation, and the socio-legal tensions emerging in generative AI. By positioning responsible AI as a sociotechnical challenge, this special day aligns with KDD's mission of advancing data science that is not only technically robust but also socially conscious.

Cybersecurity and software research have crossed paths with modern deep learning research for a few years. The power of large language models (LLMs) in particular has intrigued us to apply them to understanding binary code. In this paper, we investigate some of the many ways LLMs can be applied to binary code similarity detection, as it is a significantly more difficult task compared to source code similarity detection due to the sparsity of information and less meaningful syntax. It also has great practical implications, such as vulnerability and malware detection. We find that pretrained LLMs are mostly capable of detecting similar binary code, even with a zero-shot setting. Our main contributions and findings are to provide several supervised fine-tuning methods that, when combined, significantly surpass zero-shot LLMs and state-of-the-art binary code similarity detection methods. Specifically, we up-train the model through data augmentation, translation-style causal learning, LLM2Vec, and cumulative GTE loss. With a complete ablation study, we show that our training method can transform a generic language model into a powerful binary similarity expert, and is also robust and general enough for cross-optimization, cross-architecture, and cross-obfuscation detection.

Addressing the challenge of toxic language in online discussions is crucial for the development of effective toxicity detection models. This pioneering work focuses on addressing imbalanced datasets in toxicity detection by introducing a novel approach to augment toxic language data. We create a balanced dataset by instructing fine-tuning of Large Language Models (LLMs) using Reinforcement Learning with Human Feedback (RLHF). Recognizing the challenges in collecting sufficient toxic samples from social media platforms for building a balanced dataset, our methodology involves sentence-level text data augmentation through paraphrasing existing samples using optimized generative LLMs. Leveraging generative LLM, we utilize the Proximal Policy Optimizer (PPO) as the RL algorithm to fine-tune the model further and align it with human feedback. In other words, we start by fine-tuning a LLM using an instruction dataset, specifically tailored for the task of paraphrasing while maintaining semantic consistency. Next, we apply PPO and a reward function, to further fine-tune (optimize) the instruction-tuned LLM. This RL process guides the model in generating toxic responses. We utilize the Google Perspective API as a toxicity evaluator to assess generated responses and assign rewards/penalties accordingly. This approach guides LLMs through PPO and the reward function, transforming minority class samples into augmented versions. The primary goal of our methodology is to create a balanced and diverse dataset to enhance the accuracy and performance of classifiers in identifying instances from the minority class. Utilizing two publicly available toxic datasets, we compared various techniques with our proposed method for generating toxic samples, demonstrating that our approach outperforms all others in producing a higher number of toxic samples. Starting with an initial 16,225 toxic prompts, our method successfully generated 122,951 toxic samples with a toxicity score exceeding 30%. Subsequently, we developed various classifiers using the generated balanced datasets and applied a cost-sensitive learning approach to the original imbalanced dataset. The findings highlight the superior performance of classifiers trained on data generated using our proposed method. These results highlight the importance of employing RL and a data-agnostic model as a reward mechanism for augmenting toxic data, thereby enhancing the robustness of toxicity detection models.

Software vulnerabilities are a challenge in cybersecurity. Manual security patches are often difficult and slow to be deployed, while new vulnerabilities are created. Binary code vulnerability detection is less studied and more complex compared to source code, and this has important practical implications. Deep learning has become an efficient and powerful tool in the security domain, where it provides end-to-end and accurate prediction. Modern deep learning approaches learn the program semantics through sequence and graph neural networks, using various intermediate representation of programs, such as abstract syntax trees (AST) or control flow graphs (CFG). Due to the complex nature of program execution, the output of an execution depends on the many program states and inputs. Also, a CFG generated from static analysis can be an overestimation of the true program flow. Moreover, the size of programs often does not allow a graph neural network with fixed layers to aggregate global information. To address these issues, we propose DeepEXE, an agent-based implicit neural network that mimics the execution path of a program. We use reinforcement learning to enhance the branching decision at every program state transition and create a dynamic environment to learn the dependency between a vulnerability and certain program states. An implicitly defined neural network enables nearly infinite state transitions until convergence, which captures the structural information at a higher level. The experiments are conducted on two semi-synthetic and two real-world datasets. We show that DeepEXE is an accurate and efficient method and outperforms the state-of-the-art vulnerability detection methods.

Malware analysts often prefer reverse engineering using Call Graphs, Control Flow Graphs (CFGs), and Data Flow Graphs (DFGs), which involves the utilization of black-box Deep Learning (DL) models. The proposed research introduces a structured pipeline for reverse engineering-based analysis, offering promising results compared to state-of-the-art methods and providing high-level interpretability for malicious code blocks in subgraphs. We propose the Canonical Executable Graph (CEG) as a new representation of Portable Executable (PE) files, uniquely incorporating syntactical and semantic information into its node embeddings. At the same time, edge features capture structural aspects of PE files. This is the first work to present a PE file representation encompassing syntactical, semantic, and structural characteristics, whereas previous efforts typically focused solely on syntactic or structural properties. Furthermore, recognizing the limitations of existing graph explanation methods within Explainable Artificial Intelligence (XAI) for malware analysis, primarily due to the specificity of malicious files, we introduce Genetic Algorithm-based Graph Explainer (GAGE). GAGE operates on the CEG, striving to identify a precise subgraph relevant to predicted malware families. Through experiments and comparisons, our proposed pipeline exhibits substantial improvements in model robustness scores and discriminative power compared to the previous benchmarks. Furthermore, we have successfully used GAGE in practical applications on real-world data, producing meaningful insights and interpretability. This research offers a robust solution to enhance cybersecurity by delivering a transparent and accurate understanding of malware behaviour. Moreover, the proposed algorithm is specialized in handling graph-based data, effectively dissecting complex content and isolating influential nodes.

"This study explores the field of software reverse engineering through the lens of code summarization, which involves generating informative and concise summaries of code functionality. A significant aspect of this research is the application of assembly code summarization in malware analysis, highlighting its critical role in understanding and mitigating potential security threats. Although there have been recent efforts to develop code summarization techniques for high-level programming languages, to the best of our knowledge, this study is the first attempt to generate comments for assembly code. For this purpose, we first built a carefully curated dataset of assembly function-comment pairs. We then focused on automatic assembly code summarization using transfer learning with pre-trained natural language processing (NLP) models, including BERT, DistilBERT, RoBERTa, and CodeBERT. The results of our experiments show a notable advantage of Code- BERT: despite its initial training on high-level programming languages alone, it excels in learning assembly language, outperforming other pre-trained NLP models."@eng

Social media use has transformed communication and made social interaction more accessible. Public microblogs allow people to share and access news through existing and social-media-created social connections and access to public news sources. These benefits also create opportunities for the spread of false information. False information online can mislead people, decrease the benefits derived from social media, and reduce trust in genuine news. We divide false information into two categories: unintentional false information, also known as misinformation; and intentionally false information, also known as disinformation and fake news. Given the increasing prevalence of misinformation, it is imperative to address its dissemination on social media platforms. This survey focuses on six key aspects related to misinformation: 1) clarify the definition of misinformation to differentiate it from intentional forms of false information; 2) categorize proposed approaches to manage misinformation into three types: detection, verification, and mitigation; 3) review the platforms and languages for which these techniques have been proposed and tested; 4) describe the specific features that are considered in each category; 5) compare public datasets created to address misinformation and categorize into prelabeled content-only datasets and those including users and their connections; and 6) survey fact-checking websites that can be used to verify the accuracy of information. This survey offers a comprehensive and unprecedented review of misinformation, integrating various methodological approaches, datasets, and content-, user-, and network-based approaches, which will undoubtedly benefit future research in this field.

Object detection techniques have been widely studied, utilized in various works, and have exhibited robust performance on images with sufficient luminance. However, these approaches typically struggle to extract valuable features from low-luminance images, which often exhibit blurriness and dim appearence, leading to detection failures. To overcome this issue, we introduce an innovative unsupervised feature domain knowledge distillation (KD) framework. The proposed framework enhances the generalization capability of neural networks across both low- and high-luminance domains without incurring additional computational costs during testing. This improvement is made possible through the integration of generative adversarial networks and our proposed unsupervised KD process. Furthermore, we introduce a region-based multiscale discriminator designed to discern feature domain discrepancies at the object level rather than from the global context. This bolsters the joint learning process of object detection and feature domain distillation tasks. Both qualitative and quantitative assessments shown that the proposed method, empowered by the region-based multiscale discriminator and the unsupervised feature domain distillation process, can effectively extract beneficial features from low-luminance images, outperforming other state-of-the-art approaches in both low- and sufficient-luminance domains.

Software vulnerabilities have been posing tremendous reliability threats to the general public as well as critical infrastructures, and there have been many studies aiming to detect and mitigate software defects at the binary level. Most of the standard practices leverage both static and dynamic analysis, which have several drawbacks like heavy manual workload and high complexity. Existing deep learning-based solutions not only suffer to capture the complex relationships among different variables from raw binary code but also lack the explainability required for humans to verify, evaluate, and patch the detected bugs. We propose VulANalyzeR, a deep learning-based model, for automated binary vulnerability detection, Common Weakness Enumeration-type classification, and root cause analysis to enhance safety and security. VulANalyzeR features sequential and topological learning through recurrent units and graph convolution to simulate how a program is executed. The attention mechanism is integrated throughout the model, which shows how different instructions and the corresponding states contribute to the final classification. It also classifies the specific vulnerability type through multi-task learning as this not only provides further explanation but also allows faster patching for zero-day vulnerabilities. We show that VulANalyzeR achieves better performance for vulnerability detection over the state-of-the-art baselines. Additionally, a Common Vulnerability Exposure dataset is used to evaluate real complex vulnerabilities. We conduct case studies to show that VulANalyzeR is able to accurately identify the instructions and basic blocks that cause the vulnerability even without given any prior knowledge related to the locations during the training phase.

Representation learning has been applied to Electronic Health Records (EHR) for medical concept embedding and the downstream predictive analytics tasks with promising results. Medical ontologies can also be integrated to guide the learning so the embedding space can better align with existing medical knowledge. Yet, properly carrying out the integration is non-trivial. Medical concepts that are similar according to a medical ontology may not be necessarily close in the embedding space learned from the EHR data, as medical ontologies organize medical concepts for their own specific objectives. Any integration methodology without considering the underlying inconsistency will result in sub-optimal medical concept embedding and, in turn, degrade the performance of the downstream tasks. In this article, we propose a novel representation learning framework called ADORE ( AD aptive O ntological RE presentations) that allows the medical ontologies to adapt their structures for more robust integrating with the EHR data. ADORE first learns multiple embeddings for each category in the ontology via an attention mechanism. At the same time, it supports an adaptive integration of categorical and multi-relational ontologies in the embedding space using a category-aware graph attention network. We evaluate the performance of ADORE on a number of predictive analytics tasks using two EHR datasets. Our experimental results show that the medical concept embeddings obtained by ADORE can outperform the state-of-the-art methods for all the tasks. More importantly, it can result in clinically meaningful sub-categorization of the existing ontological categories and yield attention values that can further enhance the model interpretability.

With the increasing adoption of digital health platforms through mobile apps and online services, people have greater flexibility connecting with medical practitioners, pharmacists, and laboratories and accessing resources to manage their own health-related concerns. Many healthcare institutions are connecting with each other to facilitate the exchange of healthcare data, with the goal of effective healthcare data management. The contents generated over these platforms are often shared with third parties for a variety of purposes. However, sharing healthcare data comes with the potential risk of exposing patients’ sensitive information to privacy threats. In this article, we address the challenge of sharing healthcare data while protecting patients’ privacy. We first model a complex healthcare dataset using a heterogeneous information network that consists of multi-type entities and their relationships. We then propose DiffHetNet , an edge-based differentially private algorithm, to protect the sensitive links of patients from inbound and outbound attacks in the heterogeneous health network. We evaluate the performance of our proposed method in terms of information utility and efficiency on different types of real-life datasets that can be modeled as networks. Experimental results suggest that DiffHetNet generally yields less information loss and is significantly more efficient in terms of runtime in comparison with existing network anonymization methods. Furthermore, DiffHetNet is scalable to large network datasets.

Authorship obfuscation techniques have commonly been evaluated based on their ability to hide the author's identity (evasion) while preserving the content of the original text.However, to avoid overstating the systems' effectiveness, evasion detection must be evaluated using competitive identification techniques in settings that mimic real-life scenarios, and the outcomes of the content-preservation evaluation have to be interpretable by potential users of these obfuscation tools.Motivated by recent work on cross-topic authorship identification and content preservation in summarization, we re-evaluate different authorship obfuscation techniques on detection evasion and content preservation.Furthermore, we propose a new information-theoretic measure to characterize the misattribution harm that can be caused by detection evasion.Our results 1 reveal key weaknesses in state-of-the-art obfuscation techniques and a surprisingly competitive effectiveness from a back-translation baseline in all evaluation aspects.

Software vulnerability detection is one of the most challenging tasks faced by reverse engineers. Recently, vulnerability detection has received a lot of attention due to a drastic increase in the volume and complexity of software. Reverse engineering is a time-consuming and labor-intensive process for detecting malware and software vulnerabilities. However, with the advent of deep learning and machine learning, it has become possible for researchers to automate the process of identifying potential security breaches in software by developing more intelligent technologies. In this research, we propose VDGraph2Vec, an automated deep learning method to generate representations of assembly code for the task of vulnerability detection. Previous approaches failed to attend to topological characteristics of assembly code while discovering the weakness in the software. VDGraph2Vec embeds the control flow and semantic information of assembly code effectively using the expressive capabilities of message passing neural networks and the RoBERTa model. Our model is able to learn the important features that help distinguish between vulnerable and non-vulnerable software. We carry out our experimental analysis for performance benchmark on three of the most common weaknesses and demonstrate that our model can identify vulnerabilities with high accuracy and outperforms the current state-of-the-art binary vulnerability detection models.

Deep learning models have achieved state-of-the-art performance in many classification tasks. However, most of them cannot provide an explanation for their classification results. Machine learning models that are interpretable are usually linear or piecewise linear and yield inferior performance. Non-linear models achieve much better classification performance, but it is usually hard to explain their classification results. As a counter-example, an interpretable feedforward neural network (IFFNN) is proposed to achieve both high classification performance and interpretability for malware detection. If the IFFNN can perform well in a more flexible and general form for other classification tasks while providing meaningful explanations, it may be of great interest to the applied machine learning community. In this paper, we propose a way to generalize the interpretable feedforward neural network to multi-class classification scenarios and any type of feedforward neural networks, and evaluate its classification performance and interpretability on interpretable datasets. We conclude by finding that the generalized IFFNNs achieve comparable classification performance to their normal feedforward neural network counterparts and provide meaningful explanations. Thus, this kind of neural network architecture has great practical use.

Non-negative tensor factorization has been shown a practical solution to automatically discover phenotypes from the electronic health records (EHR) with minimal human supervision. Such methods generally require an input tensor describing the inter-modal interactions to be pre-established; however, the correspondence between different modalities (e.g., correspondence between medications and diagnoses) can often be missing in practice. Although heuristic methods can be applied to estimate them, they inevitably introduce errors, and leads to sub-optimal phenotype quality. This is particularly important for patients with complex health conditions (e.g., in critical care) as multiple diagnoses and medications are simultaneously present in the records. To alleviate this problem and discover phenotypes from EHR with unobserved inter-modal correspondence, we propose the collective hidden interaction tensor factorization (cHITF) to infer the correspondence between multiple modalities jointly with the phenotype discovery. We assume that the observed matrix for each modality is marginalization of the unobserved inter-modal correspondence, which are reconstructed by maximizing the likelihood of the observed matrices. Extensive experiments conducted on the real-world MIMIC-III dataset demonstrate that cHITF effectively infers clinically meaningful inter-modal correspondence, discovers phenotypes that are more clinically relevant and diverse, and achieves better predictive performance compared with a number of state-of-the-art computational phenotyping models.

This paper introduces a new framework for reducing mode collapse in Generative adversarial networks (GANs). The problem occurs when the generator learns to map several various input values (z) to the same output value, which makes the generator fail to capture all modes of the true data distribution. As a result, the diversity of synthetically produced data is lower than that of the real data. To address this problem, we propose a new and simple framework for training GANs based on the concept of virtual mapping. Our framework integrates two processes into GANs: merge and split. The merge process merges multiple data points (samples) into one before training the discriminator. In this way, the generator would be trained to capture the merged-data distribution rather than the (unmerged) data distribution. After the training, the split process is applied to the generator's output in order to split its contents and produce diverse modes. The proposed framework increases the chance of capturing diverse modes through enabling an indirect or virtual mapping between an input z value and multiple data points. This, in turn, enhances the chance of generating more diverse modes. Our results show the effectiveness of our framework compared to the existing approaches in terms of reducing the mode collapse problem.

Haohan Bo, Steven H. H. Ding, Benjamin C. M. Fung, Farkhund Iqbal. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

PARAFAC2 factorization provides a practical solution to map the temporally irregular electronic health records (EHR) to clinically relevant and interpretable phenotypes.Existing methods ignore the effect of interdependency of diseases over clinical history.Consequently, the crucial temporal information contained in the EHR data cannot be fully utilized and the learned phenotypes can be sub-optimal to characterize patients with progressive conditions.To address this issue, we propose a novel temporally dependent PARAFAC2 (TedPar) factorization in which the temporal dependency among the phenotypes is explicitly modeled.TedPar learns a set of target phenotypes to capture the clinical features relevant to the diseases of interest and a set of background phenotypes to capture irrelevant but frequently co-occurring clinical features.By effectively modeling the temporal dependency and separating relevant and irrelevant features, the discovered target phenotypes can be used to model the progression of the diseases of interest.Empirical evaluations show that TedPar obtains up to 32.4% relative improvement in reconstruction accuracy over the test set, suggesting significantly better generalizability than the baselines for both noise-free and heavily noisy input data.Qualitative analysis also shows that TedPar is capable of discovering clinically meaningful phenotypes and capturing the temporal dependency between them.

With the rapidly growing volume and variety of Electronic Health Records (EHR) data, deep-learning models exhibit state-of-the-art performance for many predictive tasks in the health domain. To overcome the challenge of high dimensionality in EHR data, many representation learning methods have been proposed to learn low-dimensional diagnosis representations. Another challenge is how to effectively incorporate the domain knowledge, such as the International Classification of Diseases (ICD) medical ontology, into the learned embeddings. Albeit the medical ontology is a knowledge graph, none of the existing methods take advantage of Graph Neural Network (GNN), which has demonstrated its ability in other domains. The problem is that a GNN with multiple hidden layers, which are required to propagate information from the leaf of the medical ontology graph to the root, dilutes the differences among the nodes, degrading the quality of the learned embeddings. In this paper we introduce a densely connected graph derived from the original ontology graph to tackle the problem. Furthermore, to model the information in patient records, we construct a single co-occurrence graph based on the co-occurrence of diagnoses and a patient's diagnosis history. Experimental results show that the diagnosis embeddings learned from our model, DUal-GRAph Representation Learning (DUGRA), outperform the current state-of-the-art models in terms of diagnosis prediction accuracy.

Properly analyzing health insurance claims data could lead to significant business insights and benefits for health service providers and insurance companies. Yet, health insurance data is often high dimensional and contains complex interleave sequences of claims. Instead of conducting machine learning tasks directly on the raw data, a better approach is performing the tasks on high-quality embeddings of the raw data. Driven by the real business need of Solution Segic Inc., a Canadian technology company in the group insurance industry, we extract health insurance claims embeddings with neural networks in the context of anomaly detection. We propose and thoroughly examine six embedding components that are customized based on different possible assumptions made on the data. One of our proposed embedding components, EC-ReStepRec, significantly outperforms other candidates on two anomaly detection tasks. This is the first embedding study done on health insurance claims for anomaly detection.

Non-negative Tensor Factorization (NTF) has been shown effective to discover clinically relevant and interpretable phenotypes from Electronic Health Records (EHR). Existing NTF based computational phenotyping models aggregate data over the observation window, resulting in the learned phenotypes being mixtures of disease states appearing at different times. We argue that by separating the clinical events happening at different times in the input tensor, the temporal dynamics and the disease progression within the observation window could be modeled and the learned phenotypes will correspond to more specific disease states. Yet how to construct the tensor for data samples with different temporal lengths and properly capture the temporal relationship specific to each individual data sample remains an open challenge. In this paper, we propose a novel Collective Non-negative Tensor Factorization (CNTF) model where each patient is represented by a temporal tensor, and all of the temporal tensors are factorized collectively with the phenotype definitions being shared across all patients. The proposed CNTF model is also flexible to incorporate non-temporal data modality and RNN-based temporal regularization. We validate the proposed model using MIMIC-III dataset, and the empirical results show that the learned phenotypes are clinically interpretable. Moreover, the proposed CNTF model outperforms the state-of-the-art computational phenotyping models for the mortality prediction task.

Learning representations of medical concepts from the Electronic Health Records (EHR) has been shown effective for predictive analytics in healthcare. Incorporation of medical ontologies has also been explored to further enhance the accuracy and to ensure better alignment with the known medical knowledge. Most of the existing work assumes that medical concepts under the same ontological category should share similar representations, which however does not always hold. In particular, the categorizations in medical ontologies were established with various factors being considered. Medical concepts even under the same ontological category may not follow similar occurrence patterns in the EHR data, leading to contradicting objectives for the representation learning. In this paper, we propose a deep learning model called MMORE which alleviates this conflicting objective issue by allowing multiple representations to be inferred for each ontological category via an attention mechanism. We apply MMORE to diagnosis prediction and our experimental results show that the representations obtained by MMORE can achieve better predictive accuracy and result in clinically meaningful sub-categorization of the existing ontological categories.

Reverse engineering is a manually intensive but necessary technique for understanding the inner workings of new malware, finding vulnerabilities in existing systems, and detecting patent infringements in released software. An assembly clone search engine facilitates the work of reverse engineers by identifying those duplicated or known parts. However, it is challenging to design a robust clone search engine, since there exist various compiler optimization options and code obfuscation techniques that make logically similar assembly functions appear to be very different. A practical clone search engine relies on a robust vector representation of assembly code. However, the existing clone search approaches, which rely on a manual feature engineering process to form a feature vector for an assembly function, fail to consider the relationships between features and identify those unique patterns that can statistically distinguish assembly functions. To address this problem, we propose to jointly learn the lexical semantic relationships and the vector representation of assembly functions based on assembly code. We have developed an assembly code representation learning model \emph{Asm2Vec}. It only needs assembly code as input and does not require any prior knowledge such as the correct mapping between assembly functions. It can find and incorporate rich semantic relationships among tokens appearing in assembly code. We conduct extensive experiments and benchmark the learning model with state-of-the-art static and dynamic clone search approaches. We show that the learned representation is more robust and significantly outperforms existing methods against changes introduced by obfuscation and optimizations.

Law enforcement faces problems in tracing the true identity of offenders in cybercrime investigations. Most offenders mask their true identity, impersonate people of high authority, or use identity deception and obfuscation tactics to avoid detection and traceability. To address the problem of anonymity, authorship analysis is used to identify individuals by their writing styles without knowing their actual identities. Most authorship studies are dedicated to English due to its widespread use over the Internet, but recent cyber-attacks such as the distribution of Stuxnet indicate that Internet crimes are not limited to a certain community, language, culture, ideology, or ethnicity. To effectively investigate cybercrime and to address the problem of anonymity in online communication, there is a pressing need to study authorship analysis of languages such as Arabic, Chinese, Turkish, and so on. Arabic, the focus of this study, is the fourth most widely used language on the Internet. This study investigates authorship of Arabic discourse/text, especially tiny text, Twitter posts. We benchmark the performance of a profile-based approach that uses n -grams as features and compare it with state-of-the-art instance-based classification techniques. Then we adapt an event-visualization tool that is developed for English to accommodate both Arabic and English languages and visualize the result of the attribution evidence. In addition, we investigate the relative effect of the training set, the length of tweets, and the number of authors on authorship classification accuracy. Finally, we show that diacritics have an insignificant effect on the attribution process and part-of-speech tags are less effective than character-level and word-level n -grams.

Authorship analysis (AA) is the study of unveiling the hidden properties of authors from textual data. It extracts an author's identity and sociolinguistic characteristics based on the reflected writing styles in the text. The process is essential for various areas, such as cybercrime investigation, psycholinguistics, political socialization, etc. However, most of the previous techniques critically depend on the manual feature engineering process. Consequently, the choice of feature set has been shown to be scenario- or dataset-dependent. In this paper, to mimic the human sentence composition process using a neural network approach, we propose to incorporate different categories of linguistic features into distributed representation of words in order to learn simultaneously the writing style representations based on unlabeled texts for AA. In particular, the proposed models allow topical, lexical, syntactical, and character-level feature vectors of each document to be extracted as stylometrics. We evaluate the performance of our approach on the problems of authorship characterization, authorship identification and authorship verification with the Twitter, blog, review, novel, and essay datasets. The experiments suggest that our proposed text representation outperforms the static stylometrics, dynamic n -grams, latent Dirichlet allocation, latent semantic analysis, distributed memory model of paragraph vectors, distributed bag of words version of paragraph vector, word2vec representations, and other baselines.

Non-negative tensor factorization has been shown effective for discovering phenotypes from the EHR data with minimal human supervision. In most cases, an interaction tensor of the elements in the EHR (e.g., diagnoses and medications) has to be first established before the factorization can be applied. Such correspondence information however is often missing. While different heuristics can be used to estimate the missing correspondence, any errors introduced will in turn cause inaccuracy for the subsequent phenotype discovery task. This is especially true for patients with multiple diseases diagnosed (e.g., under critical care). To alleviate this limitation, we propose the hidden interaction tensor factorization (HITF) where the diagnosis-medication correspondence and the underlying phenotypes are inferred simultaneously. We formulate it under a Poisson non-negative tensor factorization framework and learn the HITF model via maximum likelihood estimation. For performance evaluation, we applied HITF to the MIMIC III dataset. Our empirical results show that both the phenotypes and the correspondence inferred are clinically meaningful. In addition, the inferred HITF model outperforms a number of state-of-the-art methods for mortality prediction.

Topic models were proposed to detect the underlying semantic structure of large collections of text documents to facilitate the process of browsing and accessing documents with similar ideas and topics. Applying topic models to short text documents to extract meaningful topics is challenging. The problem becomes even more complicated when dealing with short and noisy micro‐posts in Twitter that are about one general topic. In such a case, the goal of applying topic models is to extract subtopics. This results in topics represented by similar sets of keywords, which in turn makes the process of topic interpretation more confusing. In this paper we propose a new method that incorporates Twitter‐LDA, WordNet, and hashtags to enhance the keyword labels that represent each topic. We emphasize the importance of different keywords to different topics based on the semantic relationships and the co‐occurrences of keywords in hashtags. We also propose a method to find the best number of topics to represent the text document collection. Experiments on two real‐life Twitter datasets on fashion suggest that our method performs better than the original Twitter‐LDA in terms of perplexity, topic coherence, and the quality of keywords for topic labeling.

This panel critically examines the state of data: how its growth and ubiquity have confronted the computer science and particularly the database community, with new challenges. These challenges require practitioners and teachers to learn new skills and engage with other disciplines in ways they had not done before. Panelists will examine the impact of the 'bigness' of data, and its importance for an ever-increasing array of applications, as well as the implications for traditional ideas of privacy and person-hood. By bringing together data specialists with those trained in social and human sciences, this panel aims to initiate discussion about the new social role the computer scientist and the database community have begun to play.

Assembly code analysis is one of the critical processes for detecting and proving software plagiarism and software patent infringements when the source code is unavailable. It is also a common practice to discover exploits and vulnerabilities in existing software. However, it is a manually intensive and time-consuming process even for experienced reverse engineers. An effective and efficient assembly code clone search engine can greatly reduce the effort of this process, since it can identify the cloned parts that have been previously analyzed. The assembly code clone search problem belongs to the field of software engineering. However, it strongly depends on practical nearest neighbor search techniques in data mining and databases. By closely collaborating with reverse engineers and Defence Research and Development Canada (DRDC), we study the concerns and challenges that make existing assembly code clone approaches not practically applicable from the perspective of data mining. We propose a new variant of LSH scheme and incorporate it with graph matching to address these challenges. We implement an integrated assembly clone search engine called Kam1n0. It is the first clone search engine that can efficiently identify the given query assembly function's subgraph clones from a large assembly code repository. Kam1n0 is built upon the Apache Spark computation framework and Cassandra-like key-value distributed storage. A deployed demo system is publicly available. Extensive experimental results suggest that Kam1n0 is accurate, efficient, and scalable for handling large volume of assembly code.

Utility mining is a new development of data mining technology. Among utility mining problems, utility mining with the itemset share framework is a hard one as no anti-monotonicity property holds with the interestingness measure. Prior works on this problem all employ a two-phase, candidate generation approach with one exception that is however inefficient and not scalable with large databases. The two-phase approach suffers from scalability issue due to the huge number of candidates. This paper proposes a novel algorithm that finds high utility patterns in a single phase without generating candidates. The novelties lie in a high utility pattern growth approach, a lookahead strategy, and a linear data structure. Concretely, our pattern growth approach is to search a reverse set enumeration tree and to prune search space by utility upper bounding. We also look ahead to identify high utility patterns without enumeration by a closure property and a singleton property. Our linear data structure enables us to compute a tight bound for powerful pruning and to directly identify high utility patterns in an efficient and scalable way, which targets the root cause with prior algorithms. Extensive experiments on sparse and dense, synthetic and real world data suggest that our algorithm is up to 1 to 3 orders of magnitude more efficient and is more scalable than the state-of-the-art algorithms.

In the last decade, several approaches concerning private data release for data mining have been proposed. Data mashup, on the other hand, has recently emerged as a mechanism for integrating data from several data providers. Fusing both techniques to generate mashup data in a distributed environment while providing privacy and utility guarantees on the output involves several challenges. That is, how to ensure that no unnecessary information is leaked to the other parties during the mashup process, how to ensure the mashup data is protected against certain privacy threats, and how to handle the high-dimensional nature of the mashup data while guaranteeing high data utility. In this paper, we present Fusion, a privacy-preserving multi-party protocol for data mashup with guaranteed LKC-privacy for the purpose of data mining. Experiments on real-life data demonstrate that the anonymous mashup data provide better data utility, the approach can handle high dimensional data, and it is scalable with respect to the data size.

The Knowledge Discovery in Databases (KDD) comprises a number of stages, such as understanding its goals, acquiring, cleaning and integrating relevant data sources, modeling the obtained data sets by means of feature extraction and selection in order to prepare them as inputs to intelligent data analysis algorithms, and finally interpreting the results of those algorithms in practice.The success of KDD depends, among many other factors, on the ability to interact with the domain experts and the availability of effective data and information sharing.The domain knowledge is crucial for the design and tuning of practically all KDD stages.Privacy-preserving is crucial to run intelligent systems without violation of individual privacy by exposing potentially sensitive person-specific information.This special issue aims at elaborating on the state of the art in the areas of privacy-aware intelligent systems and the utilization of the domain knowledge in KDD.It is inspired by two workshops held at the 20th International Symposium on Methodologies for Intelligent Systems in Macau, China, 4-7 December 2012.The issue contains six carefully extended and thoroughly peer-reviewed articles, which were originally presented at the workshops.Their topics refer to data clustering and similarity, association rule induction and multi-label classification, as well as database engines, search systems and web services.Their results can contribute to a number of areas in our everyday lives, such as medical and healthcare systems, social and scientific networks, administrative services and so on.The first paper, by Jan Rauch, is titled "Formal Framework for Data Mining with Association Rules and Domain Knowledge -Overview of an Approach".The author outlines a methodology for data mining with association rules.The framework is based on a logical, appropriately enhanced calculus of rules.It allows to describe the whole knowledge discovery process, including formulation of analytical questions, application of the proposed analytical procedure and interpretation of its results.The domain knowledge is expressed using a formalized language aimed at deriving truly relevant rules.The second paper, by Nicolas Anciaux, Danae Boutara, Benjamin Nguyen and Michalis Vazirgiannis, is titled "Limiting Data Exposure in Multi-Label Classification Processes".The authors consider decision making processes based on multi-label classification models in the area of administrative services.They continue their previous research on minimizing data required from applicants as an input to online forms.They develop an experimental platform able to transform arbitrary multi-label data sets into collection rules and to measure the gain obtained in terms of minimizing data exposure.

The Internet provides an ideal anonymous channel for concealing computer-mediated malicious activities, as the network-based origins of critical electronic textual evidence (e.g., emails, blogs, forum posts, chat logs, etc.) can be easily repudiated. Authorship attribution is the study of identifying the actual author of the given anonymous documents based on the text itself, and for decades, many linguistic stylometry and computational techniques have been extensively studied for this purpose. However, most of the previous research emphasizes promoting the authorship attribution accuracy, and few works have been done for the purpose of constructing and visualizing the evidential traits. In addition, these sophisticated techniques are difficult for cyber investigators or linguistic experts to interpret. In this article, based on the End-to-End Digital Investigation (EEDI) framework, we propose a visualizable evidence-driven approach, namely VEA, which aims at facilitating the work of cyber investigation. Our comprehensive controlled experiment and the stratified experiment on the real-life Enron email dataset demonstrate that our approach can achieve even higher accuracy than traditional methods; meanwhile, its output can be easily visualized and interpreted as evidential traits. In addition to identifying the most plausible author of a given text, our approach also estimates the confidence for the predicted result based on a given identification context and presents visualizable linguistic evidence for each candidate.

Data-as-a-Service (DaaS) mashup enables data providers to dynamically integrate their data on demand depending on consumers' requests. Utilizing DaaS mashup, however, involves some challenges. Mashing up data from multiple sources to answer a consumer's request might reveal sensitive information and thereby compromise the privacy of individuals. Moreover, data integration of arbitrary DaaS providers might not always be sufficient to answer incoming requests. In this paper, we provide a cloud-based framework for privacy-preserving DaaS mashup that enables secure collaboration between DaaS providers for the purpose of generating an anonymous dataset to support data mining. Experiments on real-life data demonstrate that our DaaS mashup framework is scalable and can efficiently and effectively satisfy the data privacy and data mining requirements specified by the DaaS providers and the data consumers.

Privacy-preserving data publishing addresses the problem of disclosing sensitive data when mining for useful information. Among the existing privacy models, ϵ-differential privacy provides one of the strongest privacy guarantees. In this paper, we address the problem of private data publishing, where different attributes for the same set of individuals are held by two parties. In particular, we present an algorithm for differentially private data release for vertically partitioned data between two parties in the semihonest adversary model. To achieve this, we first present a two-party protocol for the exponential mechanism. This protocol can be used as a subprotocol by any other algorithm that requires the exponential mechanism in a distributed setting. Furthermore, we propose a two-party algorithm that releases differentially private data in a secure way according to the definition of secure multiparty computation. Experimental results on real-life data suggest that the proposed algorithm can effectively preserve information for a data mining task.

In this paper, we consider the collaborative data publishing problem for anonymizing horizontally partitioned data at multiple data providers. We consider a new type of “insider attack” by colluding data providers who may use their own data records (a subset of the overall data) to infer the data records contributed by other data providers. The paper addresses this new threat, and makes several contributions. First, we introduce the notion of m-privacy, which guarantees that the anonymized data satisfies a given privacy constraint against any group of up to m colluding data providers. Second, we present heuristic algorithms exploiting the monotonicity of privacy constraints for efficiently checking m-privacy given a group of records. Third, we present a data provider-aware anonymization algorithm with adaptive m-privacy checking strategies to ensure high utility and m-privacy of anonymized data with efficiency. Finally, we propose secure multi-party computation protocols for collaborative data publishing with m-privacy. All protocols are extensively analyzed and their security and efficiency are formally proved. Experiments on real-life datasets suggest that our approach achieves better or comparable utility and efficiency than existing and baseline algorithms while satisfying m-privacy.

Social network data provide valuable information for companies to better understand the characteristics of their potential customers with respect to their communities. Yet, sharing social network data in its raw form raises serious privacy concerns because a successful privacy attack not only compromises the sensitive information of the target victim but also the relationship with his/her friends or even their private information. In recent years, several anonymization techniques have been proposed to solve these issues. Most of them focus on how to achieve a given privacy model but fail to preserve the data mining knowledge required for data recipients. In this paper, we propose a method to k-anonymize a social network dataset with the goal of preserving frequent sharing patterns, one of the most important kinds of knowledge required for marketing and consumer behaviour analysis. Experimental results on real-life data illustrate the trade-off between privacy and utility loss with respect to the preservation of frequent sharing patterns.

OBJECTIVE: Privacy-preserving data publishing addresses the problem of disclosing sensitive data when mining for useful information. Among existing privacy models, ε-differential privacy provides one of the strongest privacy guarantees and makes no assumptions about an adversary's background knowledge. All existing solutions that ensure ε-differential privacy handle the problem of disclosing relational and set-valued data in a privacy-preserving manner separately. In this paper, we propose an algorithm that considers both relational and set-valued data in differentially private disclosure of healthcare data. METHODS: The proposed approach makes a simple yet fundamental switch in differentially private algorithm design: instead of listing all possible records (ie, a contingency table) for noise addition, records are generalized before noise addition. The algorithm first generalizes the raw data in a probabilistic way, and then adds noise to guarantee ε-differential privacy. RESULTS: We showed that the disclosed data could be used effectively to build a decision tree induction classifier. Experimental results demonstrated that the proposed algorithm is scalable and performs better than existing solutions for classification analysis. LIMITATION: The resulting utility may degrade when the output domain size is very large, making it potentially inappropriate to generate synthetic data for large health databases. CONCLUSIONS: Unlike existing techniques, the proposed algorithm allows the disclosure of health data containing both relational and set-valued data in a differentially private manner, and can retain essential information for discriminative analysis.

In this paper, we study the problem of privacy preserving record linkage which aims to perform record linkage without revealing anything about the non-linked records. We propose a new secure embedding strategy based on frequent variable length grams which allows record linkage on the embedded space. The frequent grams used for constructing the embedding base are mined from the original database under the framework of differential privacy. Compared with the state-of-the-art secure matching schema [15], our approach provides formal, provable privacy guarantees and achieves better scalability while providing comparable utility.

Utility mining emerged recently to address the limitation of frequent itemset mining by introducing interestingness measures that reflect both the statistical significance and the user's expectation. Among utility mining problems, utility mining with the itemset share framework is a hard one as no anti-monotone property holds with the interestingness measure. The state-of-the-art works on this problem all employ a two-phase, candidate generation approach, which suffers from the scalability issue due to the huge number of candidates. This paper proposes a high utility itemset growth approach that works in a single phase without generating candidates. Our basic approach is to enumerate itemsets by prefix extensions, to prune search space by utility upper bounding, and to maintain original utility information in the mining process by a novel data structure. Such a data structure enables us to compute a tight bound for powerful pruning and to directly identify high utility itemsets in an efficient and scalable way. We further enhance the efficiency significantly by introducing recursive irrelevant item filtering with sparse data, and a lookahead strategy with dense data. Extensive experiments on sparse and dense, synthetic and real data suggest that our algorithm outperforms the state-of-the-art algorithms over one order of magnitude.

With the wide deployment of smart card automated fare collection (SCAFC) systems, public transit agencies have been benefiting from huge volume of transit data, a kind of sequential data, collected every day. Yet, improper publishing and use of transit data could jeopardize passengers' privacy. In this paper, we present our solution to transit data publication under the rigorous differential privacy model for the Société de transport de Montréal (STM). We propose an efficient data-dependent yet differentially private transit data sanitization approach based on a hybrid-granularity prefix tree structure. Moreover, as a post-processing step, we make use of the inherent consistency constraints of a prefix tree to conduct constrained inferences, which lead to better utility. Our solution not only applies to general sequential data, but also can be seamlessly extended to trajectory data. To our best knowledge, this is the first paper to introduce a practical solution for publishing large volume of sequential data under differential privacy. We examine data utility in terms of two popular data analysis tasks conducted at the STM, namely count queries and frequent sequential pattern mining. Extensive experiments on real-life STM datasets confirm that our approach maintains high utility and is scalable to large datasets.

Cyber criminals exploit opportunities for anonymity and masquerade in web-based communication to conduct illegal activities such as phishing, spamming, cyber predation, cyber threatening, blackmail, and drug trafficking. One way to fight cyber crime is to collect digital evidence from online documents and to prosecute cyber criminals in the court of law. In this paper, we propose a unified framework using data mining and natural language processing techniques to analyze online messages for the purpose of crime investigation. Our framework takes the chat log from a confiscated computer as input, extracts the social networks from the log, summarizes chat conversations into topics, identifies the information relevant to crime investigation, and visualizes the knowledge for an investigator. To ensure that the implemented framework meets the needs of law enforcement officers in real-life investigation, we closely collaborate with the cyber crime unit of a law enforcement agency in Canada. Both the feedback from the law enforcement officers and experimental results suggest that the proposed chat log mining framework is effective for crime investigation.

Mashup is a web technology that allows different service providers to flexibly integrate their expertise and to deliver highly customizable services to their customers. Data mashup is a special type of mashup application that aims at integrating data from multiple data providers depending on the user's request. However, integrating data from multiple sources brings about three challenges: 1) Simply joining multiple private data sets together would reveal the sensitive information to the other data providers. 2) The integrated (mashup) data could potentially sharpen the identification of individuals and, therefore, reveal their person-specific sensitive information that was not available before the mashup. 3) The mashup data from multiple sources often contain many data attributes. When enforcing a traditional privacy model, such as K-anonymity, the high-dimensional data would suffer from the problem known as the curse of high dimensionality, resulting in useless data for further data analysis. In this paper, we study and resolve a privacy problem in a real-life mashup application for the online advertising industry in social networks, and propose a service-oriented architecture along with a privacy-preserving data mashup algorithm to address the aforementioned challenges. Experiments on real-life data suggest that our proposed architecture and algorithm is effective for simultaneously preserving both privacy and information utility on the mashup data. To the best of our knowledge, this is the first work that integrates high-dimensional data for mashup service.

Privacy-preserving data publishing addresses the problem of disclosing sensitive data when mining for useful information. Among the existing privacy models, ∈-differential privacy provides one of the strongest privacy guarantees and has no assumptions about an adversary's background knowledge. Most of the existing solutions that ensure ∈-differential privacy are based on an interactive model, where the data miner is only allowed to pose aggregate queries to the database. In this paper, we propose the first anonymization algorithm for the non-interactive setting based on the generalization technique. The proposed solution first probabilistically generalizes the raw data and then adds noise to guarantee ∈-differential privacy. As a sample application, we show that the anonymized data can be used effectively to build a decision tree induction classifier. Experimental results demonstrate that the proposed non-interactive anonymization algorithm is scalable and performs better than the existing solutions for classification analysis.

We study the problem of anonymizing data with quasi-sensitive attributes. Quasi-sensitive attributes are not sensitive by themselves, but certain values or their combinations may be linked to external knowledge to reveal indirect sensitive information of an individual. We formalize the notion of l-diversity and t-closeness for quasi-sensitive attributes, which we call QS l-diversity and QS t-closeness, to prevent indirect sensitive attribute disclosure. We propose a two-phase anonymization algorithm that combines quasi-identifying value generalization and quasi-sensitive value suppression to achieve QS l-diversity and QS t-closeness.

Sharing healthcare data has become a vital requirement in healthcare system management; however, inappropriate sharing and usage of healthcare data could threaten patients’ privacy. In this article, we study the privacy concerns of sharing patient information between the Hong Kong Red Cross Blood Transfusion Service (BTS) and the public hospitals. We generalize their information and privacy requirements to the problems of centralized anonymization and distributed anonymization , and identify the major challenges that make traditional data anonymization methods not applicable. Furthermore, we propose a new privacy model called LKC-privacy to overcome the challenges and present two anonymization algorithms to achieve LKC-privacy in both the centralized and the distributed scenarios. Experiments on real-life data demonstrate that our anonymization algorithms can effectively retain the essential information in anonymous data for data analysis and is scalable for anonymizing large datasets.

Recently, trajectory data mining has received a lot of attention in both the industry and the academic research. In this paper, we study the privacy threats in trajectory data publishing and show that traditional anonymization methods are not applicable for trajectory data due to its challenging properties: high-dimensional, sparse, and sequential. Our primary contributions are (1) to propose a new privacy model called LKC-privacy that overcomes these challenges, and (2) to develop an efficient anonymization algorithm to achieve LKC-privacy while preserving the information utility for trajectory pattern mining.

Mashup is a web technology that combines information from more than one source into a single web application. This technique provides a new platform for different data providers to flexibly integrate their expertise and deliver highly customizable services to their customers. Nonetheless, combining data from different sources could potentially reveal person-specific sensitive information. In this paper, we study and resolve a real-life privacy problem in a data mashup application for the financial industry in Sweden, and propose a privacy-preserving data mashup (PPMashup) algorithm to securely integrate private data from different data providers, whereas the integrated data still retains the essential information for supporting general data exploration or a specific data mining task, such as classification analysis. Experiments on real-life data suggest that our proposed method is effective for simultaneously preserving both privacy and information usefulness, and is scalable for handling large volume of data.

Location-based service (LBS) is gaining momentum as GPS-equipped mobile devices become increasingly affordable and popular. One of the potential obstacles faced by LBS is that users may raise concerns about their personal privacy when location data are sent to a distrusted LBS provider. A well-known solution is to render the location data less accurate through spatial or temporal cloaking. In this paper, we show that by combining consecutive location data including speed, heading direction, and cloaked locations, an adversary can obtain more accurate estimation of the actual location. We propose a solution to prevent such inferences by cloaking speed and direction. Since the cloaking is based on estimated future locations, we devise methods for tolerating errors caused by the estimation process. We report simulation results on the tradeoff between the capability of tolerating errors and the degree of cloaking.

Mashup is a Web technology that combines information from more than one source into a single Web application.This technique provides a new platform for different data providers to flexibly integrate their expertise and deliver highly customizable services to their customers. None the-less, combining data from different sources could potentially reveal person-specific sensitive information. In this paper, we study and resolve a real-life privacy problem in a data mashup application for the financial industry in Sweden. Therefore we propose a service-oriented architecture for privacy-preserving data mashup together with a multi-party protocol to securely integrate private data from different data providers, whereas the integrated data still retains the essential information for supporting general data exploration or a specific data mining task, such as classification analysis. Experiments on real-life data suggest that our proposed method is effective for simultaneously preserving both privacy and information usefulness.

Sharing healthcare data has become a vital requirement in healthcare system management; however, inappropriate sharing and usage of healthcare data could threaten patients' privacy. In this paper, we study the privacy concerns of the blood transfusion information-sharing system between the Hong Kong Red Cross Blood Transfusion Service (BTS) and public hospitals, and identify the major challenges that make traditional data anonymization methods not applicable. Furthermore, we propose a new privacy model called LKC-privacy, together with an anonymization algorithm, to meet the privacy and information requirements in this BTS case. Experiments on the real-life data demonstrate that our anonymization algorithm can effectively retain the essential information in anonymous data for data analysis and is scalable for anonymizing large datasets.

k-anonymization is an important privacy protection mechanism in data publishing. While there has been a great deal of work in recent years, almost all considered a single static release. Such mechanisms only protect the data up to the first release or first recipient. In practical applications, data is published continuously as new data arrive; the same data may be anonymized differently for a different purpose or a different recipient. In such scenarios, even when all releases are properly k-anonymized, the anonymity of an individual may be unintentionally compromised if recipient cross-examines all the releases received or colludes with other recipients. Preventing such attacks, called correspondence attacks, faces major challenges. In this paper, we systematically characterize the correspondence attacks and propose an efficient anonymization algorithm to thwart the attacks in the model of continuous data publishing. 1.

We consider the problem of publishing sensitive transaction data with privacy preservation. High dimensionality of transaction data poses unique challenges on data privacy and data utility. On one hand, re-identification attacks tend to use a subset of items that infrequently occur in transactions, called moles. On the other hand, data mining applications typically depend on subsets of items that frequently occur in transactions, called nuggets. Thus the problem is how to eliminate all moles while retaining nuggets as much as possible. A challenge is that moles and nuggets are multi-dimensional with exponential growth and are tangled together by shared items. We present a novel and scalable solution to this problem. The novelty lies in a compact border data structure that eliminates the need of generating all moles and nuggets.

Releasing person-specific data could potentially reveal sensitive information of individuals. k-anonymization is a promising privacy protection mechanism in data publishing. Though substantial research has been conducted on k-anonymization and its extensions in recent years, few of them consider releasing data for a specific purpose of data analysis. This paper presents a practical data publishing framework for determining a generalized version of data that preserves both individual privacy and information usefulness for cluster analysis. Experiments on real-life data suggest that, by focusing on preserving cluster structure in the generalization process, the cluster quality is significantly better than the cluster quality on the generalized data without such focus. The major challenge of generalizing data for cluster analysis is the lack of class labels that could be used to guide the generalization process. Our approach converts the problem into the counterpart problem for classification analysis where class labels encode the cluster structure in the data, and presents a framework to evaluate the cluster quality on the generalized data.

Classification is a fundamental problem in data analysis. Training a classifier requires accessing a large collection of data. Releasing person-specific data, such as customer data or patient records, may pose a threat to an individual's privacy. Even after removing explicit identifying information such as Name and SSN, it is still possible to link released records back to their identities by matching some combination of nonidentifying attributes such as {Sex, Zip, Birthdate}. A useful approach to combat such linking attacks, called k-anonymization, is anonymizing the linking attributes so that at least k released records match each value combination of the linking attributes. Previous work attempted to find an optimal k-anonymization that minimizes some data distortion metric. We argue that minimizing the distortion to the training data is not relevant to the classification goal that requires extracting the structure of predication on the "future" data. In this paper, we propose a k-anonymization solution for classification. Our goal is to find a k-anonymization, not necessarily optimal in the sense of minimizing data distortion, which preserves the classification structure. We conducted intensive experiments to evaluate the impact of anonymization on the classification on future data. Experiments on real-life data show that the quality of classification can be preserved even for highly restrictive anonymity requirements

An organization makes a new release as new information become available, releases a tailored view for each data request, releases sensitive information and identifying information separately. The availability of related releases sharpens the identification of individuals by a global quasi-identifier consisting of attributes from related releases. Since it is not an option to anonymize previously released data, the current release must be anonymized to ensure that a global quasi-identifier is not effective for identification. In this paper, we study the sequential anonymization problem under this assumption. A key question is how to anonymize the current release so that it cannot be linked to previous releases yet remains useful for its own release purpose. We introduce the lossy join, a negative property in relational database design, as a way to hide the join relationship among releases, and propose a scalable and practical solution.

Releasing person-specific data in its most specific state poses a threat to individual privacy. This paper presents a practical and efficient algorithm for determining a generalized version of data that masks sensitive information and remains useful for modelling classification. The generalization of data is implemented by specializing or detailing the level of information in a top-down manner until a minimum privacy requirement is violated. This top-down specialization is natural and efficient for handling both categorical and continuous attributes. Our approach exploits the fact that data usually contains redundant structures for classification. While generalization may eliminate some structures, other structures emerge to help. Our results show that quality of classification can be preserved even for highly restrictive privacy requirements. This work has great applicability to both public and private sectors that share information for mutual benefits and productivity.

In this paper, we present a template-based privacy preservation to protect against the threats caused by data mining abilities. The problem has dual goals: preserve the information for a wanted classification analysis and limit the usefulness of unwanted sensitive inferences that may be derived from the data. Sensitive inferences are specified by a set of "privacy templates". Each template specifies the sensitive information to be protected, a set of identifying attributes, and the maximum association between the two. We show that suppressing the domain values is an effective way to eliminate sensitive inferences. For a large data set, finding an optimal suppression is hard, since it requires optimization over all suppressions. We present an approximate but scalable solution. We demonstrate the effectiveness of this approach on real life data sets.

A major challenge in document clustering is the extremely high dimensionality. For example, the vocabulary for a document set can easily be thousands of words. On the other hand, each document often contains a small fraction of words in the vocabulary. These features require special handlings. Another requirement is hierarchical clustering where clustered documents can be browsed according to the increasing specificity of topics. In this paper, we propose to use the notion of frequent itemsets, which comes from association rule mining, for document clustering. The intuition of our clustering criterion is that each cluster is identified by some common words, called frequent itemsets, for the documents in the cluster. Frequent itemsets are also used to produce a hierarchical topic tree for clusters. By focusing on frequent items, the dimensionality of the document set is drastically reduced. We show that this method outperforms best existing methods in terms of both clustering accuracy and scalability.