Foundational Models for Malware Embeddings Using Spatio-Temporal Parallel Convolutional Networks
Authors:
Dhruv Nandakumar,
Devin Quinn,
Elijah Soba,
Eunyoung Kim,
Christopher Redino,
Chris Chan,
Kevin Choi,
Abdul Rahman,
Edward Bowen
Abstract:
In today's interconnected digital landscape, the proliferation of malware poses a significant threat to the security and stability of computer networks and systems worldwide. As the complexity of malicious tactics, techniques, and procedures (TTPs) continuously grows to evade detection, so does the need for advanced methods capable of capturing and characterizing malware behavior. The current stat…
▽ More
In today's interconnected digital landscape, the proliferation of malware poses a significant threat to the security and stability of computer networks and systems worldwide. As the complexity of malicious tactics, techniques, and procedures (TTPs) continuously grows to evade detection, so does the need for advanced methods capable of capturing and characterizing malware behavior. The current state of the art in malware classification and detection uses task specific objectives; however, this method fails to generalize to other downstream tasks involving the same malware class. In this paper, the authors introduce a novel method that combines convolutional neural networks, standard graph embedding techniques, and a metric learning objective to extract meaningful information from network flow data and create strong embeddings characterizing malware behavior. These embeddings enable the development of highly accurate, efficient, and generalizable machine learning models for tasks such as malware strain classification, zero day threat detection, and closest attack type attribution as demonstrated in this paper. A shift from task specific objectives to strong embeddings will not only allow rapid iteration of cyber-threat detection models, but also allow different modalities to be introduced in the development of these models.
△ Less
Submitted 24 May, 2023;
originally announced May 2023.
