A survey of game theoretic approach for adversarial machine learning

Abstract

The field of machine learning is progressing at a faster pace than ever before. Many organizations leverage machine learning tools to extract useful information from a massive amount of data. In particular, machine learning finds its application in cybersecurity that begins to enter the age of automation. However, machine learning applications in cybersecurity face unique challenges other domains rarely do—attacks from active adversaries. Problems in areas such as intrusion detection, banking fraud detection, spam filtering, and malware detection have to face challenges of adversarial attacks that modify data so that malicious instances would evade detection by the learning systems. The adversarial learning problem naturally resembles a game between the learning system and the adversary. In such a game, both players would attempt to play their best strategies against each other while maximizing their own payoffs. To solve the game, each player would search for an optimal strategy against the opponent based on the prediction of the opponent's strategy choice. The problem becomes even more complicated in settings where the learning system may have to deal with many adversaries of unknown types. Applying game‐theoretic approach, robust learning techniques have been developed to specifically address adversarial attacks and the preliminary results are promising. In this review, we summarize these results. This article is categorized under: Technologies > Machine Learning Fundamental Concepts of Data and Knowledge > Key Design Issues in Data Mining

Perspectives on Adversarial Classification

“…Other approaches have focused on improving the game theoretic model in [6]. However, to our knowledge, none has been able to overcome the above mentioned unrealistic common knowledge assumptions, as may be seen in recent reviews [7,8], who point out the importance of this issue. As an example, [41] use a Stackelberg game in which both players know each other's payoff functions.…”

“…The subfield of classification that seeks for algorithms with robust behaviour against adversarial perturbations is known as adversarial classification (AC) and was pioneered by [6]. Stemming from their work, the prevailing paradigm when modelling the confrontation between classification systems and adversaries has been game theory, see recent reviews [7,8]. This entails well-known common knowledge hypothesis [9,10] according to which agents share information about their beliefs and preferences.…”

Perspectives on Adversarial Classification

“…Other approaches have focused on improving the game theoretic model in Dalvi et al [2004] but, to our knowledge, none has been able to overcome the unrealistic common knowledge assumptions, as may be seen in recent reviews by Biggio and Roli [2018] and Zhou et al [2018], who point out the importance of this issue. As an example, Kantarcıoglu et al [2011] use a Stackelberg game in which both players know each other payoff functions.…”

“…The subfield of statistical classification that seeks for algorithms with robust behavior against adversarial perturbations is known as adversarial classification (AC) and was pioneered by Dalvi et al [2004]. Stemming from their work, the prevailing paradigm used to model the confrontation between classification systems and adversaries has been game theory, see recent reviews by Biggio and Roli [2018] and Zhou et al [2018]. This entails well-known common knowledge hypothesis [Antos and Pfeffer 2010] according to which agents share information about their utilities and probabilities.…”

Mimicking Anti-Viruses with Machine Learning and Entropy Profiles

“…The works of Goodfellow et al [ 60 ] try to improve the capabilities of classifiers by detecting the adversarial attack directly. Other works leverage game theory, defining a reactive game between the attacker and the defender, and trying to measure its Nash equilibrium [ 61 ]. This helps to understand the limitations of the classifiers.…”