Those who commit cybercrime know they need to stay on the cutting edge of technology to come up with new and different ways to swindle people. Luckily, the good guys are also spending time in research labs developing ways to thwart the latest tricks employed by spammers, phishers and other criminals.
Below is a list of a dozen research projects underway that focus on new technology and techniques to stop spam. While in many cases these projects are reacting to exploits already in use, such as image spam and phishing, the work by these researchers is designed to counter spammers' current developments and may also lead to prevention of future ones. This list, by no means complete, contains select papers recently made public.
Spam filter makers were stumped when image spam made its debut last Spring; by hiding the spam message inside an image that filters couldn't discern, spammers got their messages through to in-boxes.
"Learning Fast Classifiers for Image Spam" is the name of a research paper from the University of Pennsylvania that describes how filters can be tweaked to quickly determine whether or not an inbound message containing an image is spam. The paper discusses techniques that focus on simple properties of the image to make classifications as fast as possible, the development of an algorithm that can select features for classification based on speed and predictive power, and a just-in-time feature extraction that "creates features at classification time as needed by the classifier," according to the paper. Researchers claim a 90% to 99% success rate using real-world data in their own tests.
Another project, "Filtering Image Spam with near-Duplicate Detection," from Princeton University, also targets spam hidden in pictures. According to the researchers behind the project, image spam is often sent in batches with visually similar images that differ only with the application of randomization algorithms. The researchers propose a near-duplicate detection system that relies on traditional antispam filtering to whittle inbound mail down to a subset of spam images, then applies multiple image-spam filters to flag all the images that look like the spam caught by traditional means. The prototype, its developers say, has reached "high detection rates" and less than 0.001% false positive (legitimate mail classified as spam) rates.
Out of Georgia Tech comes "A Discriminative Classifier Learning Approach to Image Modeling and Spam Image Identification." This proposal takes a discriminative classifier learning approach to image modeling, so that image spam can be identified. By analyzing images extracted from a body of spam messages, the researchers have identified four key image properties: color moment, color heterogeneity, conspicuousness and self-similarity. Then multiclass characterization is applied to model the images, and a maximal figure-of-merit learning algorithm is proposed to design classifiers for identifying image spam. Researchers say when tested this approach classified 81.5% of spam images correctly.
Another approach is discussed in "Image Spam Filtering by Content Obscuring Detection," from researchers at the University of Cagliari in Italy. This paper reviews low-level image processing techniques that can recognize content obscuring tricks used by spammers -- namely, character breaking and character interference via background noise -- to fool optical character recognition-detection tools.