Comparison of pattern-matching algorithms on spam email detection

Email is one of the most expedient approaches to transferring messages amongst people all over the world. Its features, in particular, its reliability, speed, and low cost makes it popular and useful among people in most areas of business and society. On the other hand, this popularity has also created new harmful actions, such as email attacks (spam) in cyberspace. Spam is arguably one of the main causes for the drowning of the World Wide Web (WWW) with many copies of similar messages generated by anonymous senders, which yields to time/space wasting of the email account holder and holds a large virus and malware threat to email providers. In spite of employing various filters to handle spam problems, such as machine learning and content-based filtering, spammers can still bypass these defense mechanisms. In this dissertation, we investigate the use of string-matching algorithms for spam email detection. In particular, this work examines and compares the efficiency of six well-known string-matching algorithms, namely the Longest Common Subsequence (LCS), the Levenshtein Distance (LD), Jaro, Jaro-Winkler, Bi-gram, and Term Frequency-Inverse Document Frequency (TFIDF) on two various datasets, which are the Enron corpus and the CSDMC2010 spam dataset. We observed that the Bi-gram algorithm performs best in spam detection; it achieved an accuracy of 99.80% and 99.85% for a variety of threshold values on Enron corpus. Moreover, it achieved a 99.95% accuracy on the CSDMC2010 dataset for all threshold values.