Abstract

Networks face many threats, ranging from DDoS attacks that overwhelm services, to compromised IoT devices, to vulnerability scanning and intrusion attempts. Many of these problems can be framed as anomalies that deviate from regular traffic, when attackers strive to blend in. Further, network managers face the perpetual need for accurate traffic forecasting to assist them in capacity planning and other traffic engineering tasks.
The AI revolution provides Machine Learning the opportunity to address long-standing problems like these that could not be resolved through deterministic algorithms. In fields such as text recognition, machine translation, image labeling, and computer games, machine learning has solved a number of long-term problems. But machine learning needs rich, diverse and huge datasets for training. Networking historically lacks such datasets in the public domain. Can we use Machine Learning to address longstanding problems in networking and cybersecurity? What are the biggest challenges to do so? We suggest that there are three barriers to overcome:(1) broadening collection and distribution of data,(2) improving and sharing of labels on that data, and (3) evaluating and developing network-specific features. Our new project, CLASSNET [10], hopes to contribute to lowering these barriers.

Date

January 1, 1970

Authors

John Heidemann, Jelena Mirkovic, Wes Hardaker, Michalis Kallitsis

Conference

NSF Workshop on AI for Networking