Shalendra Chhabra writes "Jonathan Zdziarski has been fighting spam since before the first MIT spam conference in 2003, and has now released a full-on technical book, Ending Spam, on spam filtering. Ending Spam covers how the current and near-future crop of heuristic and statistical filters actually work under the hood, and how you can most effectively use such filters to protect your inbox." Read on for the rest of Chhabra's review.
|Ending Spam: Bayesian Content Filtering and the Art of Statistical Language Classification|
|author||Jonathan A. Zdziarski|
|publisher||No Starch Press|
|summary||Very Good Book Covering Statistical Models and Techniques Implemented in Current Spam Filters|
Spam (unsolicited commercial email) and phishing (fraudulent emails) are causing losses of billions of dollars to businesses. Many initiatives are currently underway for fighting this challenge. On the legal front, a Virginia court recently sentenced a prolific spammer, Jeremy Jaynes, to nine years in prison, and a Nigerian court sentenced a woman to two and a half years for phishing. Michigan and Utah have both passed laws creating "do-not-contact" registries in July/August 2005, covering e-mail addresses, instant messaging addresses and telephone numbers. Technical initiatives to fight spam include server- or client-side spam filtering, using Lists (Blacklists, Whitelists, Greylists), Email Authentication Standards (IIM, DK, DKIM, SPF, SenderID), and emerging sender reputation and accreditation services.
Ending Spam is the first book explaining the fine details of the theoretical models and machine-learning algorithms implemented in these filters. The book is divided into three parts: introduction to spam filtering, fundamentals of statistical filtering, and advanced concepts of statistical filtering.
The first section of the book discusses the history of spam, spam kings, different approaches for fighting spam such as blacklisting, whitelisting, heuristic filtering, challenge response, throttling, collaborative filtering, Authenticated SMTP, Sender Policy Framework and SenderID, spammer fingerprinting, etc. However, the author omitted any mention of locally-sensitive hash functions (such as Nilsimsa Hash) to counter spammers' random insertion of words, the use of CAPTCHA (Completely Automated Public Turing Test to Tell Computers and Humans Apart), Greylisting, Identified Internet Mail, and Domain Keys (now Domain Keys Identified Mail).
In the next chapter, the author clearly explains various components of a Language Classifier Pipeline, including the Historical Dataset (aka wordlist, database, dictionary, filter memory), Tokenizer, and the Analysis Engine with its feedback loop. However, the process flow of a language classifier could have been more generalized, e.g. incorporating an initial text-to-text transformer. This chapter also covers the advantages and disadvantages of various training modes for filters, such as Train Everything (TEFT), Train-on-Error (TOE), and Train Until No Errors (TUNE). This part concludes with the description of Paul Graham's famous spam-filtering technique using Bayesian classification (as described in "A Plan for Spam"), Gary Robinson's Geometric Mean Test, Fisher-Robinsons Inverse Chi Square (including the source code for the inversion function), and some other tricks for optimizing spam- filtering accuracy.
The second part of this book deals with the fundamentals of statistical filtering. The author explains HTML and Base64 encoding, followed by a detailed description of tokenization techniques (e.g. Sparse Binary Polynomial Hashing). Then there's a discussion of the various tricks that spammers use for penetrating filters. Although these tactics are mentioned in John Graham-Cumming's "Spammers Compendium," Jonathan has very elegantly explained why some tricks work for spammers and some don't. This part concludes by addressing some of the resource, storage and scaling concerns raised by the large number of features generated from tokenization techniques.
The third part of this book deals with advanced concepts of statistical filtering. This includes the testing criteria for measuring accuracy of an email filter, and some advanced tokenization concepts, e.g. chained tokens (taking word-pairs and phrases into account, instead of individual words) generated using a sliding 5-byte window as mentioned in Sparse Binary Polynomial Hashing. The next chapter describes the Markovian Model implemented in the CRM114 Discriminator, but the author fails to describe different weighting schemes for features implemented in the Markovian-based version of CRM114. The author then describes the Bayesian Noise Reduction Technique for purging "out of context" data from the mail text. This chapter concludes with a very nice summary of collaborative algorithms and techniques, such as Message Innoculation, Streamlined Blackhole List, Fingerprinting, Automatic Whitelisting, URL Blacklisting, and Honeypot email addresses for snaring spammers' address harvesting bots.
The most interesting part of this book is the appendix, where the author presents interviews with John Graham-Cumming of POPFile, Brian Burton of SpamProbe, Marty Lamb of TarProxy, Bill Yerazunis of CRM114 Discriminator, and Jonathan Zdziarski of DSPAM (himself). I loved this section.
The salient points of the book: it's very easy to read; each chapter begins with a very thought-provoking introduction, and concludes with a crisp "final thoughts" section. The number of technical errors are very few in this print, and the illustrations are of good quality. Since the book is geared more toward the Bayesian and statistical generation of spam filters, the absence of certain spam-busting technologies is acceptable. However, a noticeable omission is the lack of discussion about measuring spam-filter accuracy, and what impact this has on setting filtration thresholds. A section on the economics of tradeoffs, and the use of a Receiver Operating Characteristic curve (ROC) would have been very helpful.
Overall, by putting together Ending Spam, Jonathan Zdziarski has made another significant contribution (after DSPAM) to the anti-spam community. Whether you are a system administrator, anti-spam researcher, engineer or a newbie interested in fighting spam, this book is a great reference.
William S Yerazunis and Richard Jowsey also contributed to this review. Shalendra Chhabra is a Graduate Student in Department of Computer Science and Engineering at University of California, Riverside. He is on the development team of CRM114 Discriminator and has presented his work at MIT Spam Conference 2005, Cisco Systems, and Stanford University. You can purchase Ending Spam: Bayesian Content Filtering and the Art of Statistical Language Classification from bn.com. Slashdot welcomes readers' book reviews -- to see your own review here, read the book review guidelines, then visit the submission page.