CT-T: Collaborative Research: Adaptive Attacks and Defenses in Denial of Information

点击翻译按钮获取中文摘要

Spam has become a prominent problem in every important communications medium. Most email users face spam every day. An entire industry has sprung to "improve" search engine rankings of web sites. Further, automatically generated spam has invaded blogs, social networks, online advertising, and VoIP connections. Spam is a rapidly growing practical problem due to the easy adaptation of attacking tools that bypass defense mechanisms. For example, useful defense techniques such as statistical learning filters and collaborative filtering are capable of distinguishing spam from legitimate email. However, attackers have been using automated tools to bypass these defense mechanisms, resulting in a seemingly endless "arms race" between attacks and defenses. For example, randomizing spam tokens and inserting legitimate text as camouflage can significantly reduce the effectiveness of statistical learning filters. Although there is no known general solution for the arms race, known as Adversarial Learning, a defense based on the exploitation of the semantic necessity of spam email to contain strong spam tokens such as VIAGRA (or its misspellings) has been found and demonstrated to end the camouflage arms race. This project seeks additional evidence to support the hypothesis that such structural (inherent) characteristics can be found and used in the identification of many kinds of spam attacks. The first research thrust focuses on the development of defense methods resilient to adaptive spam attacks. The second research thrust investigates the combination of spam attacks in distinct areas (e.g., email and web spam) and combined defenses. Success in these thrusts will significantly and permanently reduce the effectiveness of spam attacks.