TRUST Security Seminar: A Framework for Computing the Privacy Scores of Users in Online Social Networks
Seminar | September 30 | 1-2 p.m. | Soda Hall, Wozniak Lounge
Doug Tygar, University of California, Berkeley
Team for Research in Ubiquitous Security Technologies
This talk will survey results of the Secure Machine Learning group at
UC Berkeley. We will discuss machine learning applied to security.
Unlike conventional approaches to machine learning, security presents
Byzantine adversaries who adapt to various techniques and attempt to
make machine learning systems mislearn. We will review a number of
results:
* A taxonomy of machine learning attacks
* A successful attack on SpamBayes, a spam detector using methods
adapted from Bayesian learning
* A successful attack on a Principal Component Analysis network
anomaly detector
* A discussion of replacement network anomaly detector, ANTIDOTE,
that mitigates adversarial poisoning
* A set of results in near optimal evasion — showing results in
reverse-engineering machine learning classifiers