“Learning Using Privileged Information”

Location: LTS Auditorium, 8080 Greenmead Drive

Speaker:
Ritu Chadha
Senior Research Director, Perspecta Labs, Machine Learning and Data Analytics Research Department

Abstract:
Machine learning is widely used for a variety of activities, including cyber intrusion detection, target recognition, event prediction, and so on. Whether performed in a supervised or unsupervised manner, traditional machine learning relies on the selection of a specific set of features that are used for both training a detection model, as well as for performing the actual decision-making at runtime.

In recent work, we introduced a new machine learning paradigm called Learning Using Privileged Information (LUPI), where additional (“privileged”) features can be used to train a model, but are not required for decision-making at run-time. The use of LUPI provides several benefits, including increased accuracy, as well as a reduction in the size of training data required to achieve a certain level of decision-making accuracy.

The introduction of these “privileged” features has many applications. In the cyber arena, they allow the use of observables for training a model when those observables are not available or not suitable for use at decision time; this situation can arise if these observables can be compromised by the adversary, or if the corresponding sensors are too expensive to use at run time, or if the observables cannot be accessed at run time, etc.

We will introduce the LUPI paradigm and provide examples of its application in areas such as cybersecurity and target recognition.

Speaker Bio:
Dr. Ritu Chadha is a senior research director at Perspecta Labs, where she manages the Machine Learning and Data Analytics Research Department.

She is the PI for the ARL Cyber Security Applied Research & Experimentation Partner (AREP) 5-year, $48.5M IDIQ program and its companion ARL Cyber CRA (Collaborative Research Alliance) program, where she is leading the development of machine learning technologies applied to cyber security.

Chadha was the PI for the recently concluded DARPA Wireless Network Defense program, where Perspecta Labs developed techniques for detecting attacks on the control plane of wireless ad hoc networks.

She authored a book on policy-driven mobile ad hoc networks in 2007. Chadha is also a Telcordia Fellow.