Forget me not: Malware Classification as a Continual Learning Problem
Oct 31, 2019·
,,·
0 min read
Maria Rigaki
Elnaz Babayeva
Sebastian Garcia
Abstract
It is well known in the machine learning community that the performance of models degrades over time with the introduction of new data, getting worse in solving both new and old tasks. This is also true for security applications of machine learning, but there is less knowledge of how this happens, when, and why. In the particular case of malware classification, the increasing growth in the amount of malicious files forces the community to research machine learning models that use data more effectively. Inspired by the way humans learn, Continual Learning algorithms try to incrementally learn new tasks from data that keeps changing, without forgetting the tasks learned in the past. In this talk we will try to address the following questions: Is Continual Learning a suitable approach for security-related datasets and problems? Can Continual Learning methods learn new malware families effectively without forgetting old ones? What are the improvements in terms of speed and storage?
Date
Oct 31, 2019 8:05 PM
Event
Cybersec & AI Prague
Authors
Postdoctoral Researcher
Maria Rigaki is a post-doctoral researcher in the Department of Computer Science at the Czech Technical University in Prague. As a member of the Stratosphere Lab, she works on the security and privacy of machine learning, and on applications of AI in cyber security. Before that she spent many years as a software developer and systems architect, working on telecommunications, physical security, emergency response systems and critical infrastructure.
In her spare time Maria enjoys hacking and playing with guitars.