An evaluation of lightweight classification methods for identifying malicious URLs

Recent research has shown that it is possible to identify malicious URLs through lexical analysis of their URL structures alone. This paper intends to explore the effectiveness of these lightweight classification algorithms when working with large real world datasets including lists of malicious URLs obtained from Phishtank as well as largely filtered benign URLs obtained from proxy traffic logs. Lightweight algorithms are defined as methods by which URLs are analysed that do not use external sources of information such as WHOIS lookups, blacklist lookups and content analysis. These parameters include URL length, number of delimiters as well as the number of traversals through the directory structure and are used throughout much of the research in the paradigm of lightweight classification. Methods which include external sources of information are often called fully featured classifications and have been shown to be only slightly more effective than a purely lexical analysis when considering both false-positives and false-negatives. This distinction allows these algorithms to be run client side without the introduction of additional latency, but still providing a high level of accuracy through the use of modern techniques in training classifiers. Analysis of this type will also be useful in an incident response analysis where large numbers of URLs need to be filtered for potentially malicious URLs as an initial step in information gathering as well as end user implementations such as browser extensions which could help protect the user from following potentially malicious links. Both AROW and CW classifier update methods will be used as prototype implementations and their effectiveness will be compared to fully featured analysis results. These methods are interesting because they are able to train on any labelled data, including instances in which their prediction is correct, allowing them to build a confidence in specific lexical features. This makes it possible for them to be trained using noisy input data, making them ideal for real world applications such as link filtering and information gathering.