Adversarial Robustness for Machine Learning summarizes the recent progress on this topic and introduces popular algorithms on adversarial attack, defense and veri?cation. Sections cover adversarial attack, veri?cation and defense, mainly focusing on image classi?cation applications which are the standard benchmark considered in the adversarial robustness community. Other sections discuss adversarial examples beyond image classification, other threat models beyond testing time attack, and applications on adversarial robustness. For researchers, this book provides a thorough literature review that summarizes latest progress in the area, which can be a good reference for conducting future research. In addition, the book can also be used as a textbook for graduate courses on adversarial robustness or trustworthy machine learning. While machine learning (ML) algorithms have achieved remarkable performance in many applications, recent studies have demonstrated their lack of robustness against adversarial disturbance. The lack of robustness brings security concerns in ML models for real applications such as self-driving cars, robotics controls and healthcare systems. Summarizes the whole field of adversarial robustness for Machine learning models Provides a clearly explained, self-contained reference Introduces formulations, algorithms and intuitions Includes applications based on adversarial robustness
This book demonstrates the optimal adversarial attacks against several important signal processing algorithms. Through presenting the optimal attacks in wireless sensor networks, array signal processing, principal component analysis, etc, the authors reveal the robustness of the signal processing algorithms against adversarial attacks. Since data quality is crucial in signal processing, the adversary that can poison the data will be a significant threat to signal processing. Therefore, it is necessary and urgent to investigate the behavior of machine learning algorithms in signal processing under adversarial attacks. The authors in this book mainly examine the adversarial robustness of three commonly used machine learning algorithms in signal processing respectively: linear regression, LASSO-based feature selection, and principal component analysis (PCA). As to linear regression, the authors derive the optimal poisoning data sample and the optimal feature modifications, and also demonstrate the effectiveness of the attack against a wireless distributed learning system. The authors further extend the linear regression to LASSO-based feature selection and study the best strategy to mislead the learning system to select the wrong features. The authors find the optimal attack strategy by solving a bi-level optimization problem and also illustrate how this attack influences array signal processing and weather data analysis. In the end, the authors consider the adversarial robustness of the subspace learning problem. The authors examine the optimal modification strategy under the energy constraints to delude the PCA-based subspace learning algorithm. This book targets researchers working in machine learning, electronic information, and information theory as well as advanced-level students studying these subjects. R&D engineers who are working in machine learning, adversarial machine learning, robust machine learning, and technical consultants working on the security and robustness of machine learning are likely to purchase this book as a reference guide.
Deep Neural Networks (DNNs) have made many breakthroughs in different areas of artificial intelligence. However, recent studies show that DNNs are vulnerable to adversarial examples. A tiny perturbation on an image that is almost invisible to human eyes could mislead a well-trained image classifier towards misclassification. This raises serious security concerns and trustworthy issues towards the robustness of Deep Neural Networks in solving real world challenges. Researchers have been working on this problem for a while and it has further led to a vigorous arms race between heuristic defenses that propose ways to defend against existing attacks and newly-devised attacks that are able to penetrate such defenses. While the arm race continues, it becomes more and more crucial to accurately evaluate model robustness effectively and efficiently under different threat models and identify those ``falsely'' robust models that may give us a false sense of robustness. On the other hand, despite the fast development of various kinds of heuristic defenses, their practical robustness is still far from satisfactory, and there are actually little algorithmic improvements in terms of defenses during recent years. This suggests that there still lacks further understandings toward the fundamentals of adversarial robustness in deep learning, which might prevent us from designing more powerful defenses. \\The overarching goal of this research is to enable accurate evaluations of model robustness under different practical settings as well as to establish a deeper understanding towards other factors in the machine learning training pipeline that might affect model robustness. Specifically, we develop efficient and effective Frank-Wolfe attack algorithms under white-box and black-box settings and a hard-label adversarial attack, RayS, which is capable of detecting ``falsely'' robust models. In terms of understanding adversarial robustness, we propose to theoretically study the relationship between model robustness and data distributions, the relationship between model robustness and model architectures, as well as the relationship between model robustness and loss smoothness. The techniques proposed in this dissertation form a line of researches that deepens our understandings towards adversarial robustness and could further guide us in designing better and faster robust training methods.
Efficient operation and control of modern day urban systems such as transportation networks is now more important than ever due to huge societal benefits. Low cost network-wide sensors generate large amounts of data which needs to processed to extract useful information necessary for operational maintenance and to perform real-time control. Modern Machine Learning (ML) systems, particularly Deep Neural Networks (DNNs), provide a scalable solution to the problem of information retrieval from sensor data. Therefore, Deep Learning systems are increasingly playing an important role in day-to-day operations of our urban systems and hence cannot not be treated as standalone systems anymore. This naturally raises questions from a security viewpoint. Are modern ML systems robust to adversarial attacks for deployment in critical real-world applications? If not, then how can we make progress in securing these systems against such attacks? In this thesis we first demonstrate the vulnerability of modern ML systems on a real world scenario relevant to transportation networks by successfully attacking a commercial ML platform using a traffic-camera image. We review different methods of defense and various challenges associated in training an adversarially robust classifier. In terms of contributions, we propose and investigate a new method of defense to build adversarially robust classifiers using Error-Correcting Codes (ECCs). The idea of using Error-Correcting Codes for multi-class classification has been investigated in the past but only under nominal settings. We build upon this idea in the context of adversarial robustness of Deep Neural Networks. Following the guidelines of code-book design from literature, we formulate a discrete optimization problem to generate codebooks in a systematic manner. This optimization problem maximizes minimum hamming distance between codewords of the codebook while maintaining high column separation. Using the optimal solution of the discrete optimization problem as our codebook, we then build a (robust) multi-class classifier from that codebook. To estimate the adversarial accuracy of ECC based classifiers resulting from different codebooks, we provide methods to generate gradient based white-box attacks. We discuss estimation of class probability estimates (or scores) which are in itself useful for real-world applications along with their use in generating black-box and white-box attacks. We also discuss differentiable decoding methods, which can also be used to generate white-box attacks. We are able to outperform standard all-pairs codebook, providing evidence to the fact that compact codebooks generated using our discrete optimization approach can indeed provide high performance. Most importantly, we show that ECC based classifiers can be partially robust even without any adversarial training. We also show that this robustness is simply not a manifestation of the large network capacity of the overall classifier. Our approach can be seen as the first step towards designing classifiers which are robust by design. These contributions suggest that ECCs based approach can be useful to improve the robustness of modern ML systems and thus making urban systems more resilient to adversarial attacks.
A critical challenge in deep learning is the vulnerability of deep learning networks to security attacks from intelligent cyber adversaries. Even innocuous perturbations to the training data can be used to manipulate the behaviour of deep networks in unintended ways. In this book, we review the latest developments in adversarial attack technologies in computer vision; natural language processing; and cybersecurity with regard to multidimensional, textual and image data, sequence data, and temporal data. In turn, we assess the robustness properties of deep learning networks to produce a taxonomy of adversarial examples that characterises the security of learning systems using game theoretical adversarial deep learning algorithms. The state-of-the-art in adversarial perturbation-based privacy protection mechanisms is also reviewed. We propose new adversary types for game theoretical objectives in non-stationary computational learning environments. Proper quantification of the hypothesis set in the decision problems of our research leads to various functional problems, oracular problems, sampling tasks, and optimization problems. We also address the defence mechanisms currently available for deep learning models deployed in real-world environments. The learning theories used in these defence mechanisms concern data representations, feature manipulations, misclassifications costs, sensitivity landscapes, distributional robustness, and complexity classes of the adversarial deep learning algorithms and their applications. In closing, we propose future research directions in adversarial deep learning applications for resilient learning system design and review formalized learning assumptions concerning the attack surfaces and robustness characteristics of artificial intelligence applications so as to deconstruct the contemporary adversarial deep learning designs. Given its scope, the book will be of interest to Adversarial Machine Learning practitioners and Adversarial Artificial Intelligence researchers whose work involves the design and application of Adversarial Deep Learning.
The increasing abundance of large high-quality datasets, combined with significant technical advances over the last several decades have made machine learning into a major tool employed across a broad array of tasks including vision, language, finance, and security. However, success has been accompanied with important new challenges: many applications of machine learning are adversarial in nature. Some are adversarial because they are safety critical, such as autonomous driving. An adversary in these applications can be a malicious party aimed at causing congestion or accidents, or may even model unusual situations that expose vulnerabilities in the prediction engine. Other applications are adversarial because their task and/or the data they use are. For example, an important class of problems in security involves detection, such as malware, spam, and intrusion detection. The use of machine learning for detecting malicious entities creates an incentive among adversaries to evade detection by changing their behavior or the content of malicius objects they develop. The field of adversarial machine learning has emerged to study vulnerabilities of machine learning approaches in adversarial settings and to develop techniques to make learning robust to adversarial manipulation. This book provides a technical overview of this field. After reviewing machine learning concepts and approaches, as well as common use cases of these in adversarial settings, we present a general categorization of attacks on machine learning. We then address two major categories of attacks and associated defenses: decision-time attacks, in which an adversary changes the nature of instances seen by a learned model at the time of prediction in order to cause errors, and poisoning or training time attacks, in which the actual training dataset is maliciously modified. In our final chapter devoted to technical content, we discuss recent techniques for attacks on deep learning, as well as approaches for improving robustness of deep neural networks. We conclude with a discussion of several important issues in the area of adversarial learning that in our view warrant further research. Given the increasing interest in the area of adversarial machine learning, we hope this book provides readers with the tools necessary to successfully engage in research and practice of machine learning in adversarial settings.
Machine learning has made enormous progress in the last five to ten years. We can now make a computer, a machine, learn complex perceptual tasks from data rather than explicitly programming it. When we compare modern speech or image recognition systems to those from a decade ago, the advances are awe-inspiring. The susceptibility of machine learning systems to small, maliciously crafted adversarial perturbations is less impressive. Almost imperceptible pixel shifts or background noises can completely derail their performance. While humans are often amused by the stupidity of artificial intelligence, engineers worry about the security and safety of their machine learning applications, and scientists wonder how to make machine learning models more robust and more human-like. This dissertation summarizes and discusses advances in three areas of adversarial robustness. First, we introduce a new type of adversarial attack against machine learning models in real-world black-box scenarios. Unlike previous attacks, it does not require any insider knowledge or special access. Our results demonstrate the concrete threat caused by the current lack of robustness in machine learning applications. Second, we present several contributions to deal with the diverse challenges around evaluating adversarial robustness. The most fundamental challenge is that common attacks cannot distinguish robust models from models with misleading gradients. We help uncover and solve this problem through two new types of attacks immune to gradient masking. Misaligned incentives are another reason for insufficient evaluations. We published joint guidelines and organized an interactive competition to mitigate this problem. Finally, our open-source adversarial attacks library Foolbox empowers countless researchers to overcome common technical obstacles. Since robustness evaluations are inherently unstandardized, straightforward access to various attacks is more than a technical convenience; it promotes thorough evaluations. Third, we showcase a fundamentally new neural network architecture for robust classification. It uses a generative analysis-by-synthesis approach. We demonstrate its robustness using a digit recognition task and simultaneously reveal the limitations of prior work that uses adversarial training. Moreover, further studies have shown that our model best predicts human judgments on so-called controversial stimuli and that our approach scales to more complex datasets.
The Intelligence Community Studies Board (ICSB) of the National Academies of Sciences, Engineering, and Medicine convened a workshop on December 11â€"12, 2018, in Berkeley, California, to discuss robust machine learning algorithms and systems for the detection and mitigation of adversarial attacks and anomalies. This publication summarizes the presentations and discussions from the workshop.
Neural networks have been widely adopted to address different real-world problems. Despite the remarkable achievements in machine learning tasks, they remain vulnerable to adversarial examples that are imperceptible to humans but can mislead the state-of-the-art models. More specifically, such adversarial examples can be generalized to a variety of common data structures, including images, texts and networked data. Faced with the significant threat that adversarial attacks pose to security-critical applications, in this thesis, we explore the good, the bad and the ugly of adversarial machine learning. In particular, we focus on the investigation on the applicability of adversarial attacks in real-world scenarios for social good and their defensive paradigms. The rapid progress of adversarial attacking techniques aids us to better understand the underlying vulnerabilities of neural networks that inspires us to explore their potential usage for good purposes. In real world, social media has extremely reshaped our daily life due to their worldwide accessibility, but its data privacy also suffers from inference attacks. Based on the fact that deep neural networks are vulnerable to adversarial examples, we attempt a novel perspective of protecting data privacy in social media and design a defense framework called Adv4SG, where we introduce adversarial attacks to forge latent feature representations and mislead attribute inference attacks. Considering that text data in social media shares the most significant privacy of users, we investigate how text-space adversarial attacks can be leveraged to protect users' attributes. Specifically, we integrate social media property to advance Adv4SG, and introduce cost-effective mechanisms to expedite attribute protection over text data under the black-box setting. By conducting extensive experiments on real-world social media datasets, we show that Adv4SG is an appealing method to mitigate the inference attacks. Second, we extend our study to more complex networked data. Social network is more of a heterogeneous environment which is naturally represented as graph-structured data, maintaining rich user activities and complicated relationships among them. This enables attackers to deploy graph neural networks (GNNs) to automate attribute inferences from user features and relationships, which makes such privacy disclosure hard to avoid. To address that, we take advantage of the vulnerability of GNNs to adversarial attacks, and propose a new graph poisoning attack, called AttrOBF to mislead GNNs into misclassification and thus protect personal attribute privacy against GNN-based inference attacks on social networks. AttrOBF provides a more practical formulation through obfuscating optimal training user attribute values for real-world social graphs. Our results demonstrate the promising potential of applying adversarial attacks to attribute protection on social graphs. Third, we introduce a watermarking-based defense strategy against adversarial attacks on deep neural networks. With the ever-increasing arms race between defenses and attacks, most existing defense methods ignore fact that attackers can possibly detect and reproduce the differentiable model, which leaves the window for evolving attacks to adaptively evade the defense. Based on this observation, we propose a defense mechanism that creates a knowledge gap between attackers and defenders by imposing a secret watermarking process into standard deep neural networks. We analyze the experimental results of a wide range of watermarking algorithms in our defense method against state-of-the-art attacks on baseline image datasets, and validate the effectiveness our method in protesting adversarial examples. Our research expands the investigation of enhancing the deep learning model robustness against adversarial attacks and unveil the insights of applying adversary for social good. We design Adv4SG and AttrOBF to take advantage of the superiority of adversarial attacking techniques to protect the social media user's privacy on the basis of discrete textual data and networked data, respectively. Both of them can be realized under the practical black-box setting. We also provide the first attempt at utilizing digital watermark to increase model's randomness that suppresses attacker's capability. Through our evaluation, we validate their effectiveness and demonstrate their promising value in real-world use.
Deep learning based classification algorithms perform poorly on adversarially perturbed data. Adversarial risk quantifies the performance of a classifier in the presence of an adversary. Numerous definitions of adversarial risk---not all mathematically rigorous and differing subtly in the details---have appeared in the literature. Adversarial attacks are designed to increase the adversarial risk of classifiers, and robust classifiers are sought that can resist such attacks. It was hitherto unknown what the theoretical limits on adversarial risk are, and whether there is an equilibrium in the game between the classifier and the adversary. In this thesis, we establish a mathematically rigorous foundation for adversarial robustness, derive algorithm-independent bounds on adversarial risk, and provide alternative characterizations based on distributional robustness and game theory. Key to these results are the numerous connections we discover between adversarial robustness and optimal transport theory. We begin by examining various definitions for adversarial risk, and laying down conditions for their measurability and equivalences. In binary classification with 0-1 loss, we show that the optimal adversarial risk is determined by an optimal transport cost between the probability distributions of the two classes. Using the couplings that achieve this cost, we derive the optimal robust classifiers for several univariate distributions. Using our results, we compute lower bounds on adversarial risk for several real-world datasets. We extend our results to general loss functions under convexity and smoothness assumptions. We close with alternative characterizations for adversarial robustness that lead to the proof of a pure Nash equilibrium in the two-player game between the adversary and the classifier. We show that adversarial risk is identical to the minimax risk in a robust hypothesis testing problem with Wasserstein uncertainty sets. Moreover, the optimal adversarial risk is the Bayes error between a worst-case pair of distributions belonging to these sets. Our theoretical results lead to several algorithmic insights for practitioners and motivate further study on the intersection of adversarial robustness and optimal transport.
