By Seetha Debi Antalan

The internet is swarmed with a lot of data traffic. With this, it becomes difficult to look for patterns of security breach attempts and other changes from normal network behavior. In attempting to more effectively address the risk of data theft and similar cybersecurity threats or concerns, researchers are turning to Machine Learning (ML) techniques.

In this approach, a large amount of data is fed to a high performance computer for training and validation. The machine will “learn” what an anomalous activity is and what is not. It could then attempt to track what is neither immediately nor usually visible to human analysts. The results will be based on the data given to the machine so selecting and engineering the input data is crucial. The suitability of the data representation as well as the complexity and depth of models are key factors in ensuring the success of the machine learning algorithm.

A team of researchers from the DOST-ASTI and UP-EEEI documented these findings in a technical paper, “Efficient Feature Extraction for Internet Data Analysis Using AS2Vec”, published in the Proceedings of the 33rd Annual ACM Symposium on Applied Computing.

The paper proposes a more efficient and effective way to generate pattern results through the use of Artificial Intelligence (AI) techniques on suitably engineered and selected raw data. As described in the study, machine learning algorithms require a feature vector that describes an individual or element. The goal is to find certain patterns from these features and subsequently make predictions based on these patterns. Derivation of these features, however, is both resource and labor-intensive and requires ready access to domain expertise and large amount of measured network data.

They proposed a feature extraction process that generates highly-relevant numerical features for representing Autonomous Systems (AS) in analyzing Internet routing data. When analyzed, these extracted features contained distinctive patterns relevant to the AS. Using these features in machine learning, they were able to accurately predict an imbalanced AS classification problem which likely indicating an anomalous condition.

The researchers tackled problems in deriving relevant features of Autonomous Systems (AS) on the Internet.