Practical Evasion of a Learning-Based Classifier: A Case Study

Abstract

Abstract-Learning-based classifiers are increasingly used for detection of various forms of malicious data. However, if they are deployed online, an attacker may attempt to evade them by manipulating the data. Examples of such attacks have been previously studied under the assumption that an attacker has full knowledge about the deployed classifier. In practice, such assumptions rarely hold, especially for systems deployed online. A significant amount of information about a deployed classifier system can be obtained from various sources. In this paper, we experimentally investigate the effectiveness of classifier evasion using a real, deployed system, PDFRATE, as a test case. We develop a taxonomy for practical evasion strategies and adapt known evasion algorithms to implement specific scenarios in our taxonomy. Our experimental results reveal a substantial drop of PDFRATE's classification scores and detection accuracy after it is exposed even to simple attacks. We further study potential defense mechanisms against classifier evasion. Our experiments reveal that the original technique proposed for PDFRATE is only effective if the executed attack exactly matches the anticipated one. In the discussion of the findings of our study, we analyze some potential techniques for increasing robustness of learningbased systems against adversarial manipulation of data.

Fortifying Botnet Classification based on Venn-abers Prediction

“…Wagner and Soto [14] demonstrated the mimicry attack against a host-based IDS that mimics the legitimate sequence of system calls. Srndic and Laskov [15] presented a mimicry attack against PDF Rate [16], a system to detect malicious pdf files based on the random forest classifier.…”

“…Srndic and Laskov [9] applied a gradient descent-kernel density estimation attack against the PDF Rate system that uses SVM and random forest classifier. Biggio et al [10] demonstrated a gradient descent component against the SVM classifier and a neural network.…”

“…And over 70% of the advanced malware created today uses one or more evasion techniques to avoid detection 2 . Attackers collect the knowledge of machine learning based detection approach and develop new evasion techniques, such as new communication channel [4][5][6][7][8], mimicry attack [9,10], gradient descent attack [9,10], poison attack [11], and so on.…”

Adversarial Active Learning

“…Drift refers to the non-stationarity of data where the data distribution changes over time. This drift can be natural gradual drift, for example, changes to a user's preference over time, or adversarial drift where an adversary changes the data to purposefully decrease the classification accuracy [14,25,53]. For example, using malware re-packaging toolkits, known as Fully Un-Detectable or FUD crypters, malware vendors repackage malware to evade anti-virus tools [7].…”

On the Practicality of Integrity Attacks on Document-Level Sentiment Analysis

“…For instance, an attacker inserts positively labeled datapoints with words of heavy negative sentiment.To our knowledge, our work is the first to demonstrate adversarial machine learning attacks in such a practical scenario. Recent work has also considered practical scenarios, but focused on evasion attacks [20,21].…”