DeLTA seminar by Nirupam Gupta: Machine Learning in Untrusted Environment
Speaker
Nirupam Gupta, postdoc at Computer Science Department at EPFL.
Title
Machine Learning in Untrusted Environment
Abstract
Recent AI systems make use of machine-learning algorithms, where the computing system learns from data (observations) in order to adjust its behavior. In fact, the availability of huge amounts of data has been the key enabler of modern AI-based technologies. This dependency on data, however, is also the Achille's heel of AI systems. The data, which can come from a wide variety of sources, is not always trustworthy. Some sources can provide erroneous or corrupted data. With current machine-learning algorithms, a single "bad" source can lead the entire learning scheme to make critical mistakes. Moreover, to handle the huge amounts of data, machine-learning algorithms are often deployed over a large number of computing machines. Consequently, as this number increases, the likelihood of machine errors also increases. Indeed, software and hardware bugs are prevalent. Furthermore, machines can sometimes be hacked by malicious players, either directly or indirectly through viruses. Some of these players attempt to corrupt the entire learning procedure, merely for the pleasure of claiming to have destroyed an important system. Others attempt to influence the learning procedure for their own benefit.
Building machine-learning schemes that are robust to these events is paramount to transitioning AI from being a mere spectacle capable of momentary feats to a dependable tool with guaranteed safety. In this talk, we cover some effective techniques for achieving such robustness. We present machine-learning algorithms that do not trust any individual data source or computing unit. We do assume, however, that a majority of data sources and machines are trustworthy; otherwise, no meaningful learning guarantee can be provided. The challenge arises from the fact that the identity of the trusted data sources and machines is a priori unknown.
_____________________________
You can subscribe to the DeLTA Seminar mailing list by sending an empty email to delta-seminar-join@list.ku.dk.
Online calendar
DeLTA Lab page
DeLTA is a research group affiliated with the Department of Computer Science at the University of Copenhagen studying diverse aspects of Machine Learning Theory and its applications, including, but not limited to Reinforcement Learning, Online Learning and Bandits, PAC-Bayesian analysis