It takes 3 pixels to make the seeing blind

Datum

11.07.2019

Zeit

14:30 - 16:00

Sprecher

Dr. Peter Steinbach

Zugehörigkeit

Max Planck Institute of Molecular Cell Biology and Genetics

Serie

IMB - Seminar

Sprache

en

Hauptthema

Biologie

Andere Themen

Medizin

Host

IMB

Beschreibung

Deep Learning based image classification has become ubiquitous in many fields. However, industrial and academic stakeholders have found ways to trick a trained network in order to force a deliberate misclassification. The adversarial examples created by these methods are in-distinguishable by eye sight from their originals and yet cause standard deep learning networks to misclassify the input. This is called an adversarial attack. In this talk, I'll introduce adversarial attacks in the field of medical imaging. I will discuss the nomenclature and mathematical tricks behind the scenes. Finally, I will show results of a black-box attack on a image classification model that was trained to differentiate healthy from sick humans with Diabetic Retinopathy based on an open dataset. The talk is concluded with a discussion of why adversarial examples exist.

Letztmalig verändert: 03.06.2019, 12:07:15