Cyber Security Webinar by Jan-Philipp Schulze MSc – From Adversarial Examples to Point Anomalies: Detecting Anomalous Inputs by Analysing the Gradient of Neural Networks

Meeting details

Meeting link: Click here

Meeting ID: 953 9411 2249

Passcode: 600391

Abstract

Neural networks learn by a gradient-based optimisation scheme. During training, the model weights are adapted to the classes seen in the training set. In our work, we analyse the differences in the gradient distribution, leading to our adversarial example detection method DA3G and our anomaly detection method R2-AD2.

Instead of hand-crafted features in the gradient space, e.g. magnitudes, alignments or similarities, we analyse the raw gradient of several layers. Thanks to our data-driven architecture, our detection methods are readily applicable to several fields of application - and may be a valuable building block for your next research project.

Short bio:

As a research fellow at Fraunhofer AISEC, a leading security research institute located in Munich, Germany, Jan-Philipp Schulze focuses on the intersection of machine learning and cybersecurity. He pursues a PhD in IT security at the Technical University of Munich.
 
His research covers anomaly detection and the robustness of deep learning methods with his work published at top conferences like KDD, ECML-PKDD and ESORICS. Prior to his time in Munich, he studied electrical engineering and information technology at ETH Zurich, in which he holds a BSc and MSc degree.