Abstract
Traditional email security focuses on scanning incoming messages, which are often too noisy to establish a reliable behavioral baseline. This paper proposes a shift toward profiling outgoing email traffic to detect self-propagating worms at the source. By extracting two dozen behavioral features and applying Maximum Covariance Analysis for feature selection, the authors show that pruning redundant features significantly improves the accuracy of SVM and Naïve Bayes classifiers. Evaluation on the Enron corpus and real-world worm traces demonstrates the system's ability to detect zero-day threats by identifying deviations from established user behavioral clusters.
Introduction
Traditional anti-virus and spam filters primarily analyze incoming emails using static signatures or statistical text analysis. However, an incoming mailbox is noisy and contains messages from many distinct users, making it difficult to establish a baseline of "normal" behavior for the mailbox owner.
Furthermore, when a machine is compromised by a self-propagating worm (like MyDoom or Netsky), it turns the host into a spam/malware cannon. The authors argue that by profiling an individual's outgoing email traffic, defenders can establish a behavioral baseline and rapidly detect abnormal activity at the source. This allows for the containment of novel, zero-day worms before security vendors even have time to generate and distribute new hash signatures.
Methodology and Feature Selection
The authors define two dozen behavioral features divided into per-email statistics (e.g., presence of HTML, binary attachments) and window-based trends (e.g., sending rate, unique recipients). To identify the most discriminative features, they utilize Maximum Covariance Analysis. Unlike standard PCA, which only maximizes variance, this technique finds directions in the feature space that maximize the covariance between the observations and their actual labels (normal vs. viral).
A key insight from this analysis is that users naturally group into common behavioral clusters. While building a unique model for every user is data-intensive, covariance analysis reveals that a single model can effectively represent a large set of similar users, making deployment at scale feasible.
Key Results: The Goldilocks Zone
The top-ranked features for detecting worm behavior included the ratio of emails with attachments, the presence of binary attachments (specifically application/octet-stream), and the frequency of emails sent in a given window.
Both the One-Class SVM and Naïve Bayes classifiers showed distinct performance curves based on feature count. Accuracy peaked when using an optimally pruned subset of features and degraded significantly when the model was bloated with irrelevant or redundant ones. This "Goldilocks Zone" suggests that too few features lead to high false positives, while too many introduce noise and cause overfitting.
The system successfully detected novel (untrained) worms in the test set by recognizing the behavioral anomaly of the outgoing traffic, proving its viability as a zero-day containment mechanism.
Conclusion and Future Work
While effective, the greedy feature selection method still struggles with redundancy; for instance, multiple top features may flag the same underlying phenomenon, such as executable attachments. Future work will focus on more rigorous pruning techniques, such as Fisher discriminant analysis, and incorporating social network analysis to refine the behavioral models. Real-world deployment will also require periodic retraining to account for temporal changes in user behavior.
Summary updated