An image processing approach to traffic anomaly detection

摘要

This paper discusses the possibility of applying an image-processing technique to detecting anomalies in Internet traffic, which is different from traditional techniques of detecting anomalies. We first demonstrate that anomalous packet behavior in darknet traces often has a characteristic multi-scale structure in time and space (e.g., in addresses or ports). These observed structures consist of abnormal and non random uses of particular traffic features. From the observations, we propose a new type of algorithm for detecting anomalies based on a technique of pattern recognition. The key idea underlying our algorithm is that anomalous activities appear as "lines" on temporal-spatial planes, which are easily identified by an edge-detection algorithm. Also, the application of a clustering technique to the lines obtained helps in classifying and labeling the numerous anomalies detected. The proposed algorithm was used to blindly analyze packet traffic traces collected from a trans-Pacific transit link. Furthermore, we compared the anomalies detected by our algorithm with those found by a statistical-based algorithm. Consequently, the comparison revealed that the two algorithms found mainly the same anomalies but some were of various different characteristic types.