The goal is to "identify the source of illegitimate email that has been anonymized", or falsely attributed to someone else.

If a specific author cannot be identified, it would also be useful to be able to infer some profile information about the author, such as gender, if english was a first language, education, or age.

The goal is to determine the author of a given text, given a set of past texts from all possible authors that are properly attributed to those authors.  This is done by examining the structural and linguistic features of a text rather than the metadata (such as From: address, which can be spoofed).

A "pattern of authorship" will be built from each of $n$ possible authors by studying texts from these $n$ authors.  Then will apply tools such as those available in LEMUR (toolkit for language modeling and information retrieval (IR)).  THese tools will be used to build a model of authorship for several models.  THese will then be used to determine the auhor of an unknown email.

Some outlying questions to be studyed are:
*how long messages have to be to be able to determine authorship with any distinguishing probability. (>1000 words?)
*how many messages are needed to build a pattern of authorship.  What timeframe must they come from (an author's style changes with time)?  Subject matter?
*reduce the number of suspected authors

Also want to identify stylometric features that can be used to determine authorship.  How is email different from other texts?  We will specifically study a dataset of newsgroup posts collected from various authors and study our techniques on that dataset.

POssible distinguishing features of text that can be used to model authors are word length distributions, sentence length distributions, function word frequencies, email structural features, freqency of 2-grams (a sequence of two letters in text.), macro-structural features of text (greetings, signatures, and their presence/absense) etc.  Some features that may not help determine authorship are context-dependant word frequencies and subject of text.

The field of stylometry assumes that authors may have ingrained (and subconcious) writing styles that are characterized by "core word usage, sentence complexity", and punctuation style used.  These features are used to define an author's style, and statistical methods are applied to determine differences between two authors.  

This research has many applications, from determining the author of literary works to determining the source of anonymous email threats.  THe Federalist papers have been studyied in the former category (There are 12 disputed papers, written either by Madison or Hamilton).

We begin our research by studying Mosteller and Wallace's book on authorship of the Federalist Papers.  The 85 Federalist papers were written in the years 1787-1788 to persuade the poeple of New York to adopt the consitution.  Of these papers, 5 are attributed to John Jay, 51 to Alexander Hamilton, 14 to Madison, 3 jointly to the latter two authors, and a remaining 12 are disputed between Hamilton and Madison.  Mosteller and Wallace used Bayesian analysis to build models of authorship for each author, and conluded that the 12 disputed papers were written by Madison.  After we study the techniques used in this book, as well as read a few more more recent works on email authorship identification (Corney thesis), we will begin our analysis of the email dataset mentioned previously.