ITU / Research / PhD Programme / Courses / Archive / 2022 / August / PhD Symposium - Summer School on Privacy-Preserving Machine Learning

PhD Symposium - Summer School on Privacy-Preserving Machine Learning

August 1 - August 4

For more information and sign-up, please visit: https://www.conferencemanager.dk/ppmlschool2022

Organizers:

Bernardo David (ITU)
Carsten Baum (AU)

Lecturer:

Divya Gupta (Senior Researcher at Microsoft, India)
Yuriy Polyakov (Senior Researcher at Duality, Israel)
Ilya Mironov (Senior Researcher at Meta, USA)
Peter Kairouz (Senior Researcher at Google, USA)
Yang Zhang (Research Group Leader at CISPA, Germany)
Rafael Dowsley (Ass. Prof. at Monash University, Australia)

Dates of the course:

01/08/2022 - 04/08/2022

Time:

9:00-17:00

Room:

Auditorium 3

Course description:

Our goal is to attract around 60-80 attendants both from theory and practice, in particular highly talented and motivated PhD and Master students. Our school will cover techniques for privacy preserving machine learning such as Secure Computation for Machine Learning, Federated Learning, Differential Privacy & Attacks such as Model Extraction. The choice of speakers which we are inviting reflects this variety of topics.

Reading list

Reading list:

Secure Computation:

“An Introduction to Secret-Sharing-Based Secure Multiparty Computation” by Daniel Escudero, https://eprint.iacr.org/2022/062.pdf
“A Pragmatic Introduction to Secure Multiparty Computation” by David Evans, Vlad Kolesnikov and Mike Rosulek, https://securecomputation.org/docs/pragmaticmpc.pdf

Fully-Homomorphic Encryption:

“A Guide to Fully Homomorphic Encryption”, Frederik Armknecht et al., https://eprint.iacr.org/2015/1192.pdf

Adversarial Machine Learning:

“Stealing Machine Learning Models via Prediction APIs”, Florian Tramer et al., https://www.usenix.org/system/files/conference/usenixsecurity16/sec16_paper_tramer.pdf
“Intriguing properties of neural networks”, Christian Szegedy et al., https://arxiv.org/abs/1312.6199

Federated Learning and Differential Privacy:

“Advances and Open Problems in Federated Learning”, Peter Kairouz et al. , https://arxiv.org/pdf/1912.04977.pdf
“Deep Learning with Differential Privacy”, Martin Abadi et al., https://dl.acm.org/doi/abs/10.1145/2976749.2978318

Programme

Programme:

Each day will have lectures at 9:00-12:00 with a coffee break at 10:30-11:00, lunch at 12:30-13:30, lectures at 13:30-17:00 with a coffee break at 15:00-15:30.

Day 1:
- PPML from general-purpose MPC (morning, Dr. Gupta) & special-purpose protocols (afternoon, Prof. Dowsley)
Day 2:
- FHE for PPML (morning, Dr. Polyakov) & Adversarial examples and backdoors (afternoon, Prof. Zhang)
Day 3:
- Differential Privacy & Federated Learning (morning, Dr. Kairouz), free afternoon and School dinner
Day 4:
- Federated Learning (morning, Dr. Mironov), Model extraction and similar attacks (afternoon, Prof. Zhang)

Prerequisites:

Students are expected to have finished a Bachelors degree in either Computer Science or Mathematics and therefore have an understanding of Linear Algebra and elementary Number and Probability Theory.

Further, students are expected to have finished at least an introductory course into Machine Learning and be familiar with standard algorithms and training techniques. Finally, students preferably have taken a course in basic data privacy, cryptography and/or cryptographic protocols ahead of the school.

Exam:

The examined student will, after the school, submit a report in which the student must discuss and reflect upon a recent publication in the area of Privacy-Preserving Machine Learning. The publication will be chosen differently per student by the organizers. The student will put the results from the given paper into perspective of the state-of-the-art of the area, as taught during the school, and will discuss how it improves upon it. We further expect the student to outline ideas for how to improve the result further. The student is then asked to submit a report, of at most 10 pages, until August 30th, to both Bernardo David and Carsten Baum. The student will be graded as pass/fail based on this report.

Credits:

3 ECTS

Amount of hours the student is expected to use on the course:

Preparation for course (studying reading material): 22.5
Participation: 30
Report preparation: 30

Participants:

We plan to recruit around 60-80 excellent Master as well as PhD students both domestically and from abroad.

How to sign up:

Students are expected to register providing their name, email address, affiliation, and supervisor name via the summer school website (https://www.conferencemanager.dk/ppmlschool2022). The registration will only incur a small fee of at most 500 DKK.