Department of Computer Science | Institute of Theoretical Computer Science | CADMO

Theory of Combinatorial Algorithms

Prof. Emo Welzl and Prof. Bernd Gärtner

Mittagsseminar (in cooperation with A. Steger, D. Steurer and B. Sudakov)

Mittagsseminar Talk Information

Date and Time: Tuesday, June 18, 2019, 12:15 pm

Duration: 30 minutes

Location: CAB G 11

Speaker: Justin Lewis (UT Austin)

Provable Certificates for Adversarial Examples: Fitting a Ball in the Union of Polytopes

This talk will cover the theory and operation of a geometrically inspired method for certifying the robustness of Piecewise Linear Neural Networks to adversarial attacks. Specifically, the problem of interest is to find the largest sized lp ball around an input image x0 such that the output class of the network within this region remains unchanged. In the case of PLNNs, it can be shown that this problem amounts to finding the largest, centered lp ball which fits within a union of polytopes. Additionally, it can be shown that these polytopes are members of a polyhedral complex and thus finding such a largest sized ball within their union can be done in time polynomial in the relevant dimensions. [Key-topics: Adversarial Examples | Deep Learning Theory | Safety in Machine Learning]

Upcoming talks     |     All previous talks     |     Talks by speaker     |     Upcoming talks in iCal format (beta version!)

Previous talks by year:   2023  2022  2021  2020  2019  2018  2017  2016  2015  2014  2013  2012  2011  2010  2009  2008  2007  2006  2005  2004  2003  2002  2001  2000  1999  1998  1997  1996  

Information for students and suggested topics for student talks

Automatic MiSe System Software Version 1.4803M   |   admin login