EE Colloquium: Revealing Backdoors, Post-Training in Deep Neural Network Classifiers

Prof. David Miller, Penn State, will be the speaker.

Abstract:

This talk covers Dr. Miller’s research on detecting backdoor data poisoning attacks on deep neural network classifiers. Here, the classifier learns from poisoned data containing a backdoor pattern that may either be imperceptible or innocuous/scene plausible. Defenses against these attacks proposed by Dr. Miller and his research group, which are state-of-the-art and require no access to the data set used to train the classifier, will be highlighted. Adversarial learning, including the problem of backdoor data poisoning, is central to machine learning in general, as it reveals existing weaknesses of machine learning/deep learning, which need to be overcome for deep learning to realize its (promised) application potential.

Bio:

David J. Miller is professor in the School of EECS at Penn State, where he has been faculty since 1995. His research interests encompass many problems in machine learning, as well as in statistical signal processing. He is also co-founder of the startup company Anomalee, Inc.