This paper studies enhancement of training data D to improve the robustness of machine learning (ML) classifiers M against adversarial attacks on relational data. Data enhancing aims to (a) defuse poisoned imperceptible features embedded in D , and (b) defend against attacks at prediction time that are unseen in D . We show that while there exists an inherent tradeoff between the accuracy and robustness of M in case (b), data enhancing can improve both the accuracy and robustness at the same time in case (a). We formulate two data enhancing problems accordingly, and show that both problems are intractable.Despite the hardness, we propose a framework that integrates model training and data enhancing. Moreover, we develop algorithms for (a) detecting and debugging corrupted imperceptible features in training data, and (b) selecting and adding adversarial examples to training data to defend against unseen attacks at prediction time. Using real-life datasets, we empirically verify that the method is at least 20.4% more robust and 2.02X faster than SOTA methods for classifiers M , without degrading the accuracy of M .

There has been a host of work on entity resolution (ER), to identify tuples that refer to the same entity. This paper studies the inverse of ER, to identify tuples to which distinct real-world entities are matched by mistake, and split such tuples into a set of tuples, one for each entity. We formulate the tuple splitting problem. We propose a scheme to decide what tuples to split and what tuples to correct without splitting, fix errors/assign attribute values to the split tuples, and impute missing values. The scheme introduces a class of rules, which embed predicates for aligning entities across relations and knowledge graphs G, assessing correlation between attributes, and extracting data from G. It unifies logic deduction, correlation models, and data extraction by chasing the data with the rules. We train machine learning models to assess attribute correlation and predict missing values. We develop algorithms for the tuple splitting scheme. Using real-life data, we empirically verify that the scheme is efficient and accurate, with F-measure 0.92 on average.

In a wide spectrum of real-world applications, it is very important to analyze and mine graph data such as social networks, communication networks, citation networks, and so on. However, the release of such graph data often raises privacy issue, and the graph privacy preservation has recently drawn much attention from the database community. While prior works on graph privacy preservation mainly focused on protecting the privacy of either the graph structure only or vertex attributes only, in this paper, we propose a novel mechanism for graph privacy preservation by considering attacks from both graph structures and vertex attributes, which transforms the original graph to a so-called$kt$-safe graph, via$k$-anonymity and$t$-closeness. We prove that the generation of a$kt$-safe graph is NP-hard, therefore, we propose a feasible framework for effectively and efficiently anonymizing a graph with low anonymization cost. In particular, we design a cost-model-based graph partitioning approach to enable our proposed divide-and-conquer strategy for the graph anonymization, and propose effective optimization techniques such as pruning method and a tree synopsis to improve the anonymization efficiency over large-scale graphs. Extensive experiments have been conducted to verify the efficiency and effectiveness of our proposed$kt$-safe graph generation approach on both real and synthetic data sets.

In many real applications such as the data integration, social network analysis, and the Semantic Web, the entity resolution (ER) is an important and fundamental problem, which identifies and links the same real-world entities from various data sources. While prior works usually consider ER over static and complete data, in practice, application data are usually collected in a streaming fashion, and often incur missing attributes (due to the inaccuracy of data extraction techniques). Therefore, in this paper, we will formulate and tackle a novel problem, topic-aware entity resolution over incomplete data streams (TER-iDS), which online imputes incomplete tuples and detects pairs of topic-related matching entities from incomplete data streams. In order to effectively and efficiently tackle the TER-iDS problem, we propose an effective imputation strategy, carefully design effective pruning strategies, as well as indexes/synopsis, and develop an efficient TER-iDS algorithm via index joins. Extensive experiments have been conducted to evaluate the effectiveness and efficiency of our proposed TER-iDS approach over real data sets.

For decades, the join operator over fast data streams has always drawn much attention from the database community, due to its wide spectrum of real-world applications, such as online clustering, intrusion detection, sensor data monitoring, and so on. Existing works usually assume that the underlying streams to be joined are complete (without any missing values). However, this assumption may not always hold, since objects from streams may contain some missing attributes, due to various reasons such as packet losses, network congestion/failure, and so on. In this paper, we formalize an important problem, namely join over incomplete data streams (Join-iDS), which retrieves joining object pairs from incomplete data streams with high confidences. We tackle the Join-iDS problem in the style of "data imputation and query processing at the same time". To enable this style, we design an effective and efficient cost-model-based imputation method via deferential dependency (DD), devise effective pruning strategies to reduce the Join-iDS search space, and propose efficient algorithms via our proposed cost-model-based data synopsis/indexes. Extensive experiments have been conducted to verify the efficiency and effectiveness of our proposed Join-iDS approach on both real and synthetic data sets.