Max-Planck-Institut für Informatik
max planck institut
informatik
mpii logo Minerva of the Max Planck Society
 

MPI-INF or MPI-SWS or Local Campus Event Calendar

<< Previous Entry Next Entry >> New Event Entry Edit this Entry Login to DB (to update, delete)
What and Who
Title:Image Manipulation against Learned Models: Privacy and Security Implications
Speaker:Seong Joon Oh
coming from:Max-Planck-Institut für Informatik - D2
Speakers Bio:
Event Type:Promotionskolloquium
Visibility:D1, D2, D3, D4, D5, RG1, SWS, MMCI
We use this to send out email in the morning.
Level:Public Audience
Language:English
Date, Time and Location
Date:Monday, 6 August 2018
Time:16:00
Duration:60 Minutes
Location:Saarbrücken
Building:E1 4
Room:024
Abstract
Machine learning is transforming the world. Its application areas span privacy sensitive and security critical tasks such as human identification and self-driving cars. These applications raise privacy and security related questions that are not fully understood or answered yet: Can automatic person recognisers identify people in photos even when their faces are blurred? How easy is it to find an adversarial input for a self-driving car that makes it drive off the road?

This thesis contributes one of the first steps towards a better understanding of such concerns in the presence of data manipulation. From the point of view of user's privacy, we show the inefficacy of common obfuscation methods like face blurring, and propose more advanced techniques based on head inpainting and adversarial examples. We discuss the duality of model security and user privacy problems and describe the implications of research in one area for the other. Finally, we study the knowledge aspect of the data manipulation problem: the more one knows about the target model, the more effective manipulations one can craft. We propose a game theoretic framework to systematically represent the partial knowledge on the target model and derive privacy and security guarantees. We also demonstrate that one can reveal architectural details and training hyperparameters of a model only by querying it, leading to even more effective data manipulations against it.

Contact
Name(s):Connie Balzert
Phone:2000
EMail:cbalzert@mpi-inf.mpg.de
Video Broadcast
Video Broadcast:NoTo Location:
Tags, Category, Keywords and additional notes
Note:
Attachments, File(s):

Created:Connie Balzert/MPI-INF, 06/27/2018 09:25 AM Last modified:Uwe Brahm/MPII/DE, 06/29/2018 07:01 AM
  • Connie Balzert, 06/27/2018 09:25 AM -- Created document.