Biblio

Machine learning and data mining algorithms typically assume that the training and testing data are sampled from the same fixed probability distribution; however, this violation is often violated in practice. The field of domain adaptation addresses the situation where this assumption of a fixed probability between the two domains is violated; however, the difference between the two domains (training/source and testing/target) may not be known a priori. There has been a recent thrust in addressing the problem of learning in the presence of an adversary, which we formulate as a problem of domain adaption to build a more robust classifier. This is because the overall security of classifiers and their preprocessing stages have been called into question with the recent findings of adversaries in a learning setting. Adversarial training (and testing) data pose a serious threat to scenarios where an attacker has the opportunity to ``poison'' the training or ``evade'' on the testing data set(s) in order to achieve something that is not in the best interest of the classifier. Recent work has begun to show the impact of adversarial data on several classifiers; however, the impact of the adversary on aspects related to preprocessing of data (i.e., dimensionality reduction or feature selection) has widely been ignored in the revamp of adversarial learning research. Furthermore, variable selection, which is a vital component to any data analysis, has been shown to be particularly susceptible under an attacker that has knowledge of the task. In this work, we explore avenues for learning resilient classification models in the adversarial learning setting by considering the effects of adversarial data and how to mitigate its effects through optimization. Our model forms a single convex optimization problem that uses the labeled training data from the source domain and known- weaknesses of the model for an adversarial component. We benchmark the proposed approach on synthetic data and show the trade-off between classification accuracy and skew-insensitive statistics.

Machine learning is enabling a myriad innovations, including new algorithms for cancer diagnosis and self-driving cars. The broad use of machine learning makes it important to understand the extent to which machine-learning algorithms are subject to attack, particularly when used in applications where physical security or safety is at risk. In this paper, we focus on facial biometric systems, which are widely used in surveillance and access control. We define and investigate a novel class of attacks: attacks that are physically realizable and inconspicuous, and allow an attacker to evade recognition or impersonate another individual. We develop a systematic method to automatically generate such attacks, which are realized through printing a pair of eyeglass frames. When worn by the attacker whose image is supplied to a state-of-the-art face-recognition algorithm, the eyeglasses allow her to evade being recognized or to impersonate another individual. Our investigation focuses on white-box face-recognition systems, but we also demonstrate how similar techniques can be used in black-box scenarios, as well as to avoid face detection.

Machine learning is widely used in security-sensitive settings like spam and malware detection, although it has been shown that malicious data can be carefully modified at test time to evade detection. To overcome this limitation, adversary-aware learning algorithms have been developed, exploiting robust optimization and game-theoretical models to incorporate knowledge of potential adversarial data manipulations into the learning algorithm. Despite these techniques have been shown to be effective in some adversarial learning tasks, their adoption in practice is hindered by different factors, including the difficulty of meeting specific theoretical requirements, the complexity of implementation, and scalability issues, in terms of computational time and space required during training. In this work, we aim to develop secure kernel machines against evasion attacks that are not computationally more demanding than their non-secure counterparts. In particular, leveraging recent work on robustness and regularization, we show that the security of a linear classifier can be drastically improved by selecting a proper regularizer, depending on the kind of evasion attack, as well as unbalancing the cost of classification errors. We then discuss the security of nonlinear kernel machines, and show that a proper choice of the kernel function is crucial. We also show that unbalancing the cost of classification errors and varying some kernel parameters can further improve classifier security, yielding decision functions that better enclose the legitimate data. Our results on spam and PDF malware detection corroborate our analysis.

Today's systems produce a rapidly exploding amount of data, and the data further derives more data, forming a complex data propagation network that we call the data's lineage. There are many reasons that users want systems to forget certain data including its lineage. From a privacy perspective, users who become concerned with new privacy risks of a system often want the system to forget their data and lineage. From a security perspective, if an attacker pollutes an anomaly detector by injecting manually crafted data into the training data set, the detector must forget the injected data to regain security. From a usability perspective, a user can remove noise and incorrect entries so that a recommendation engine gives useful recommendations. Therefore, we envision forgetting systems, capable of forgetting certain data and their lineages, completely and quickly. This paper focuses on making learning systems forget, the process of which we call machine unlearning, or simply unlearning. We present a general, efficient unlearning approach by transforming learning algorithms used by a system into a summation form. To forget a training data sample, our approach simply updates a small number of summations – asymptotically faster than retraining from scratch. Our approach is general, because the summation form is from the statistical query learning in which many machine learning algorithms can be implemented. Our approach also applies to all stages of machine learning, including feature selection and modeling. Our evaluation, on four diverse learning systems and real-world workloads, shows that our approach is general, effective, fast, and easy to use.