In the last decade, researchers, practitioners and companies struggled for devising mechanisms to detect cyber-security threats. Among others, those efforts originated rule-based, signature-based or supervised Machine Learning (ML) algorithms that were proven effective for detecting those intrusions that have already been encountered and characterized. Instead, new unknown threats, often referred to as zero-day attacks or zero-days, likely go undetected as they are often misclassified by those techniques. In recent years, unsupervised anomaly detection algorithms showed potential to detect zero-days. However, dedicated support for quantitative analyses of unsupervised anomaly detection algorithms is still scarce and often does not promote meta-learning, which has potential to improve classification performance. To such extent, this paper introduces the problem of zero-days and reviews unsupervised algorithms for their detection. Then, the paper applies a question-answer approach to identify typical issues in conducting quantitative analyses for zero-days detection, and shows how to setup and exercise unsupervised algorithms with appropriate tooling. Using a very recent attack dataset, we debate on i) the impact of features on the detection performance of unsupervised algorithms, ii) the relevant metrics to evaluate intrusion detectors, iii) means to compare multiple unsupervised algorithms, iv) the application of meta-learning to reduce misclassifications. Ultimately, v) we measure detection performance of unsupervised anomaly detection algorithms with respect to zero-days. Overall, the paper exemplifies how to practically orchestrate and apply an appropriate methodology, process and tool, providing even non-experts with means to select appropriate strategies to deal with zero-days.

Anomaly detection aims at identifying unexpected fluctuations in the expected behavior of a given system. It is acknowledged as a reliable answer to the identification of zero-day attacks to such extent, several ML algorithms that suit for binary classification have been proposed throughout years. However, the experimental comparison of a wide pool of unsupervised algorithms for anomaly-based intrusion detection against a comprehensive set of attacks datasets was not investigated yet. To fill such gap, we exercise 17 unsupervised anomaly detection algorithms on 11 attack datasets. Results allow elaborating on a wide range of arguments, from the behavior of the individual algorithm to the suitability of the datasets to anomaly detection. We conclude that algorithms as Isolation Forests, One-Class Support Vector Machines, and Self-Organizing Maps are more effective than their counterparts for intrusion detection, while clustering algorithms represent a good alternative due to their low computational complexity. Further, we detail how attacks with unstable, distributed, or non-repeatable behavior such as Fuzzing, Worms, and Botnets are more difficult to detect. Ultimately, we digress on capabilities of algorithms in detecting anomalies generated by a wide pool of unknown attacks, showing that achieved metric scores do not vary with respect to identifying single attacks.