by Ong Yun Qi,
Wee Kim Wee School of Communication Studies
22 August
2019
What would happen if
business centres such as Wall Street go dark for a week?"
The question was posed to
the audience at Nanyang Technopreneurship Centre's (NTC) first ever TECH TALK
on 22 August 2019. Professor Ruby B. Lee made an appearance as the event's
inaugural speaker, discussing the detection of anomalous behaviour in the power
grid with deep learning.
Professor Lee is the Forest
G Hamrick Professor in Engineering and Professor of Electrical Engineering at
Princeton University. Her research revolves around three fields in particular -
cyber security, computer architecture and deep learning - all of which were
highlighted in her presentation.
During the session, she
drew attention to security issues that the power grid system faces and the
potential threat towards national security and economic activity. Given the
power grid's role as the foundation for all computer systems and much of our
daily lives, it is a prime target for attackers.
"But the focus on
security critical infrastructure is not as well studied by the general research
community", she said.
In response, she proposed
deep learning solutions. Participants were given the opportunity to go through
her findings and the thought process behind her research.
"Our vision is a bit
broader than just protecting power grid systems," she said. "We are interested
in asking the question - is there a simple way to just continuously monitor the
security health of critical infrastructure?".
According to Professor Lee,
deep learning models are more "accurate, precise and had better
recall" than one-class machine learning models. Despite its appeal, they
come with their own set of challenges. They have little attack data to draw on,
complex processes and interdependencies to consider and they had to operate
without domain-specific knowledge.
Professor Lee then
presented her solution to the audience, a combination of the Temporal Deep
Learning (Long Short Term Model) and Reconstruction Error Distribution. Instead
of attack data, it utilises normal behaviour data for training. It compares it
with run-time data and amplifies even the most subtle deviation from the norm,
if any.
The model could also be
applicable across cyber-physical systems. As they "generally do the same
thing all the time", it allows for easier predictability of normal
behaviour.
She wrapped up her talk by
suggesting future improvements that can be made to this technology. Firstly, it
should incorporate more normal and attack data for validity. Secondly, it
should be taught to differentiate between various anomaly types. Thirdly, it
should be generalised to other systems. Lastly, it should be able to update its
reference normal data on a more frequent basis.
The session concluded with
a quick Q&A with the attendees, and Professor Lee was then presented with a
gift of appreciation.
Aimed at raising awareness
on cutting-edge technologies, TECH TALK is a series of 45-minute
lectures by thought leaders and experts from academia. It serves to allow
participants to explore new opportunities and build industry connections.
Stay tuned to our social
media and website for future updates on TECH TALK and more.
