The possibility for public attributes to disclose private information has caused widespread concern. Traditional privacy-preserving approaches have two limitations: 1) Approaches based on data anonymization or distortion often lead to poor utility-privacy trade-offs, and 2) approaches based on data encryption face heavy computational costs. These problems have prompted calls for an effective privacy-preserving framework that provides adequate privacy guarantees while maintaining good data utility. Inspired by denoising autoencoders, in this paper, we regard the information about privacy attributes contained in the public attributes as a kind of noise and design an ex ante privacy-preserving model called the Mutual Information Autoencoder (MIAE), which reconstructs the loss function of the original autoencoder by combining reconstruction errors and mutual information, and we introduce a trade-off coefficient to achieve utility-privacy trade-offs. To elucidate the superiority of the proposed model, we consider utility-privacy trade-offs with the expected distortion function as a metric of data utility and the joint mutual information as a metric of privacy disclosure, and then, we construct a convex optimization problem with multiple constraints based on rate-distortion theory. From an information theory perspective, we provide a lower bound for privacy disclosure with utility guarantees. Elaborate experiments over a real-world dataset reveal that as the level of expected distortion increases, the achievable bound obtained by MIAE exhibits a trend similar to that of the information-theoretic bound. When the expected distortion surpasses 2.2, the achievable bound obtained by MIAE also converges to 0, and the maximum gap between the achievable bound obtained by MIAE and the information-theoretic bound is no more than 1.4. Compared to existing models, MIAE can provide a tighter achievable bound and achieve good utility-privacy trade-offs.

Local differential privacy (LDP) is a prevalent measure of privacy protection as it provides rigorous privacy guarantees and has been widely studied for statistical analysis, especially in frequency estimation. As a representative LDP-enabled frequency estimation algorithm, Google'sRandomized Aggregation Privacy-Preserving Ordinal Response(RAPPOR) has been put into practice. However, it achieves sub-optimal utility due to the following limitations. Firstly, the adoption of the MD5 hash function inevitably results in the hash collision. Secondly, the application of the randomized response technique leads to randomness. To improve the practical effectiveness of RAPPOR and the utility of frequency-based services, we propose an LDP-enabled frequency estimation method called PK-RAPPOR, in which we devise an effective re-encoding hash function (RE-HF) incorporating prior knowledge (PK) about the rough frequency ranking of items. RE-HF divides items into several cohorts based on the PK and generates a unique hash value set for each item. Compared with the original RAPPOR, the hash collision can be eliminated for items from different cohorts, and the effect of randomness can be decreased by the overlapping of items from the same cohorts. We validate our proposed method with theoretical analysis and demonstrate its effectiveness with experiments on both synthetic and real-world datasets.

Anomaly detection relies on individuals' behavior profiling and works by detecting any deviation from the norm. When used for online banking fraud detection, however, it mainly suffers from three disadvantages. First, for an individual, the historical behavior data are often too limited to profile his/her behavior pattern. Second, due to the heterogeneous nature of transaction data, there lacks a uniform treatment of different kinds of attribute values, which becomes a potential barrier for model development and further usage. Third, the transaction data are highly skewed, and it becomes a challenge to utilize the label information effectively. The three disadvantages result in both poor generalization and high false positive rate of anomaly detection, and we propose a ranking metric embedding based multi-contextual behavior profiling (ReMEMBeR) model to battle them effectively. We solve the original fraud detection problem as a pseudo-recommender system problem, where an individual is treated as a pseudo-user, his/her behavior as a pseudo-item, and the label as the corresponding pseudo-rating. With the idea of collaborative filtering, for an individual, information from other similar individuals can be used to establish his/her behavior profile. In order to obtain a uniform treatment of heterogeneous attributes, we turn to an embedding based method to learn both attribute embedding and individuals' behavior profiles within a common latent space simultaneously. To utilize the label information better, our model is designed to fit pseudo-users' correct preference ranking for pseudo-items. By doing so, it explicitly learns to tell the fraudulent from the legitimate. Last but not least, we propose to identify and distinguish individuals under different contexts and further generalize the behavior profiling model to be a multi-contextual one. The proposed model can, thus, integrate the multi-contextual behavior patterns and allow transactions to be examined under the different contexts. Extensive experiments on a real-world online banking transaction dataset demonstrate that our model not only outperforms benchmarks on all metrics but also can be combined with them to achieve even better performance.

The rapid development of internet finance has caused increasing concern in online payment fraud due to its great threat. It is typical to employ rule systems or machine learning-based techniques to detect frauds. For the most significant features of such fraudulent transactions are exhibited in a sequential form, the sliding time window is a widely-recognized effective tool for this problem. With a sliding time window, features about the transaction characteristics can be extracted, and the latent patterns hidden in transaction records can be captured. However, the adaptive setting of sliding time window is really a big challenge, since the transaction patterns in real-life application scenarios are often too elusive to be captured. As a matter of fact, the practical setting usually needs to be updated and refined with manual intervention regularly. This is time-consuming indeed. In this article, we pursue an adaptive learning approach to detect fraudulent online payment transactions with automatic sliding time windows. Accordingly, we make efforts on optimizing the setting of windows and improving the adaptability. We design an intelligent window, called learning automatic window (LAW). It utilizes the learning automata to learn the proper parameters of time windows and adjust them dynamically and regularly according to the variation and oscillation of fraudulent transaction patterns. By the experiments over a real-world dataset of the online payment service from a commercial bank, we validate the gain of LAW in terms of detection effectiveness and robustness. To the best of our knowledge, this is the first work to make a sliding time window for fraud detection capable of learning its proper size in changing situations.

We aim at exploiting users' coarse behavioral records for identity theft detection in online services. We concentrate on this issue in online social networks (OSNs) that users' behavioral records usually consist of multiple dimensional behavior data. The behavioral records in each dimension are possibly coarse and insufficient for effectively modeling users' behavioral patterns. In this paper, we investigate whether there is a complementary effect among different dimensions of records for modeling users' behavioral patterns. We focus on three typical dimensions of behaviors in OSNs, i.e., offline check-ins, online tip-postings, and social contacts. We devise the dedicated behavior models based on each dimension of data, i.e., users' behavioral projection models. Then, by examining all feasible logical combinations of them, we find the optimal ones for two real-world data sets: Foursquare and Yelp. Notably, we analyze the potential correlation between customized demand and optimal logical fusion scheme. As an insightful result, we find that the correlation is independent of the specific data. This study would give the cybersecurity community new insights into the possibility and methodology to achieve a customized identity theft detection in OSNs by integrating multiple behavioral projection models.

In this paper, we aim to model the formation of interest-driven data dissemination in online social networks (OSNs). We focus on a usual type of interest-driven social sessions in OSNs, called Social-InterestCast, under which a user will autonomously determine whether to view the content from his followees depending on his interest. To figure out the formation mechanism of such a Social-InterestCast, we propose a four-layered system model, consisting of physical layer, social layer, content layer, and session layer to model this interestdriven sessions. To the best of our knowledge, this is the first work to model data dissemination in OSNs with the interestdriven characteristics.

The potential of web services comes from web service composition, where a series of single web services are chosen and composed to perform compound functionalities. As to Web service composition technology, the key step is function-level Web service discovery, which decides the classes of Web services chosen for the final purpose. This paper proposes a proto model for function-Level Web service discovery. Firstly, the idea is accepted that functions in an application category are limited and therefore can be listed and decomposed in a category function tree. Given users' function requirements for composition, it is possible to locate it exactly on the tree and decide how this function can be decomposed downside the tree to sub-functions, this procedure continues until it reaches the leaf-nodes of the tree whose functions can be described by abstract service classes. Finally, an application example is given to demonstrate the whole idea discussed.