Context learning for security
Abstract
Deep neural networks (DNNs) are being increasingly used in several sensitive applications such as surveillance, autonomous vehicles, malware classification, and network intrusion detection. However, the robustness of DNNs and their effectiveness have raised significant concerns; for example, there has been a surge in adversarial example attacks against them. Currently most DNNs do not account for a rich variety of information that is available in the input. For example, in images, they do not recognize if an object is out of place in a scene or in relation to other objects in that scene; similarly, inappropriate values in packet fields, or across a train of packets, are not accounted for by ML based network intrusion detection systems (NIDS). In this project we propose to develop robust ML models that learn relationships across the components of an input (e.g., objects in an image, or header fields in a packet), which we refer to as context, and check for consistencies when making inferences relating to security. We believe, as supported by our preliminary experiments, that context can significantly harden ML models, and also improve their effectiveness in security applications. First, the project will make significant leaps beyond our initial endeavors to consider multiple visual sensors and multi-modal sensing (e.g., visual and LiDAR). Extracting context from multiple sources, is expected to significantly improve the fidelity of the inferences, and increase robustness to adversarial subversion. Second, a deep exploration into what types of context will be effective, and what can be easily defeated by attackers will be carried out, via developing an understanding of the attack surface projected by context-based defenses, and a principled approach to tune and evaluate different forms of context. Finally, the lessons learned in the visual and sensor spaces, will be carried over to the cyber-space, where the structure of the inputs, and the constraints imposed in terms of what types of context can be extracted, make the problem more challenging, but has promise in the development of secure ML based solutions to cyber-security. The project includes an extensive validation plan involving multiple datasets (including multi-sensor/modal sets, network intrusion detection and malware datasets) and emulations on an in house camera networks vehicles and a large scale software defined network testbed. The proposed work is expected to significantly enhance the army capabilities in terms of robust surveillance and security, especially given that future ISR and command and control applications will heavily deploy machine learning. The research in the attack space, can also provide the army with an understanding of threats, and also gain the first mover advantage in tactical cyber- warfare.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 20, 2022
- Source ID
- W911NF2210260
Entities
People
- Srikanth Krishnamurthy
Organizations
- Army Contracting Command
- United States Army
- University of California, Riverside