David M. Sommer

I am a PhD student in the System Security Group at ETH Zurich. I am interested in anonymous communication, privacy, and machine learning. Recently, I focused on differential privacy and its real-world implications.

Scientific Publications

Learning Numeric Optimal Differentially Private Truncated Additive Mechanisms.

David Sommer, Lukas Abfalterer, Sheila Zingg, Esfandiar Mohammadi

arXiv preprint

arXiv

Towards Probabilistic Verification of Machine Unlearning.

David Sommer*, Liwei Song*, Sameer Wagh, Prateek Mittal

arXiv preprint

arXiv | source-code Github -- local

Cyber-Risks in Paper-Voting

David Sommer, Moritz Schneider, Jannik Gut, Srdjan Capkun.

arXiv preprint

arXiv

Privacy Loss Classes: The Central Limit Theorem in Differential Privacy

David Sommer, Sebastian Meiser, and Esfandiar Mohammadi.

Proceedings on Privacy Enhancing Technologies 2 (PoPETS), 2019

pdf | eprint | bib | DOI | slides (pptx) | slides (pdf)

Deniable Upload and Download via Passive Participation

David Sommer, Aritra Dhar, Esfandiar Mohammadi, Daniel Ronzani, and Srdjan Capkun

USENIX Symposium on Networked Systems Design and Implementation (NSDI'19), 2019

pdf | bib | DOI | slides | source-code Github -- local

Teaching Authentication in High Schools: Challenges and Lessons Learned.

Elizabeth Stobert, Elizabeta Cavar, Luka Malisa, and David Sommer

USENIX Workshop on Advances in Security Education (ASE'17), 2017

pdf

ROTE: Rollback Protection for Trusted Execution.

Sinisa Matetic, Mansoor Ahmed, Kari Kostiainen, Aritra Dhar, David Sommer, Arthur Gervais, Ari Juels, Srdjan Capkun

26th USENIX Security Symposium, 2017

pdf

Hacking in the Blind: (Almost) Invisible Runtime User Interface Attacks.

Luka Malisa, Kari Kostiainen, Thomas Knell, David Sommer, Srdjan Capkun

Conference on Cryptographic Hardware and Embedded Systems (CHES), 2017

pdf

Industry Experience

July 2020 - continuing

Privacy Consultant for Sedimentum AG, Zurich

Sedimentum AG provides a solution to inform clinical care-takers about dangerous incidents of their stationized patients while complying with the strict medical privacy law of Switzerland. By alarming care-takers, e.g., in case of falling patients, help can be provided almost instantly and eases the burden on medical professinals. I counsel Sedimentum AG in the development of their privacy preserving incident detection and alarming protocol.

March 2020 - July 2020

Privacy ML Team, Apple Cambridge UK, United Kindom

During this internship, I developed a generic framework to collect statistics of federated training data differentially private. The data we collect statistics over does never leave the devices of the users it was generated by. Compatible with the concept of federated machine learning, this knowledge allows to pick a more suitable machine learning model than when a model architecture has to be guessed blindly because there are no data samples available.

Talks

January 2020, Princeton University, Princeton, USA

The Privacy Loss Distribution and its Privacy Loss Class: The Central Limit Theorem in Differential Privacy and Other Insights

July 2019, PETS'19, Stockholm, Sweden

Privacy Loss Classes: The Central Limit Theorem in Differential Privacy

slides

Feb 2019, NSDI'19, Bosten, USA

Deniable Upload and Download via Passive Participation

slides

Nov 2017, ZISC Lunch Seminar, Zürich, Switzerland

CoverUp: Privacy Through “Forced” Participation in Anonymous Communication Networks

slides

Non-academic talks are not listed.

Teaching

WS 2020

Informatik für Mathematiker und Physiker (C++ Introduction) (TA)

WS 2019

Informatik 1 (C++ Introduction) (TA)

SS 2019

Introduction to Machine Learning (TA)

WS 2018

Informatik für Mathematiker und Physiker (C++ Introduction) (TA)

SS 2018

Design of Digital Circuits (TA)

WS 2017

Informatik für Mathematiker und Physiker (C++ Introduction) (TA)

SS 2017

Design of Digital Circuits (TA)

WS 2016

Informatik für Mathematiker und Physiker (C++ Introduction) (TA)

SS 2016

Design of Digital Circuits (TA)

Some TA slides are here