2400088 – Secure Multipary Computation

In the setting of secure multiparty computation, two or more parties with private inputs wish to compute some joint function of their inputs. The security requirements of such a computation are privacy (meaning that the parties learn the output and nothing more), correctness (meaning that the output is correctly distributed), independence of inputs, and more. Due to its generality, secure computation is a central tool in cryptography. In this seminar, we examine modern protocols for secure multiparty computation of arbitrary functions.

Allgemeine Informationen

Kursprogramm

IKNP OT-Extension
The goal of this presentation is to first motivate the idea behind extending OTs and then explain the OT-extension protocol of Ishai, Kilian, Nissim and Petrank.

Optimizing Garbled Circuits
In this presentation, various improvements on the basic garbled circuit protocol are explained:
the free-XOR optimization
the 2-row-reduction
FleXor
half-gates
(if time allows) Three Halves Make a Whole

Verifiable Secret Sharing and BGW with Active Security
In this presentation, the BGW protocol is extended to provide active security.
To this end, it is first explained what verifiable secret sharing is, and then the actively secure variant of the BGW protocol is presented.

The IPS Compiler
In this presentation, the IPS compiler to achieve active security against up to n−1 corruptions (with n parties) is presented.
The compiler combines a protocol that has passive security against n−1 corruptions and a protocol with active security for a honest majority.

The SPDZ Protocol
The goal of this presentation is to explain the SPDZ (pronounced "Speedz") protocol, a highly efficient protocol based on arithmetic secret sharing preprocessing model, where input-independent preprocessing takes place in an offline-phase with computational security, while an information-theoretic online-phase enables fast evaluation.
To this end, first the requirements towards the offline phase and the resulting precomputed values are presented, and then the online phase is explained.
Finally, if time allows, a short overview of the preprocessing phase is given.

Federated Machine Learning
Several use cases for machine learning involve data that needs to be protected, such as medical data from several different stakeholders.
Federated machine learning, an attractive framework for the massively distributed multiparty training of deep learning models, employs secure aggregation to protect participants' local models and data but now face the difficulty of detecting anomalies in participants' contribution. To federated learning(FL), the invisible private data from participants' end can be the hidden danger as well, since malicious participants get chance to manipulate remote aggregator by backdooring it through model poisoning attack.
Starting from this, following three points are expected to be explained:
1. Analyse of Vulnerability against Poison Attack in FL
2. Implementation of Backdoor Attacks in FL
3. Potential Defense against Backdoors in FL

Veranstaltungsdaten

Dozent(en)

Prof. Jörn Müller-Quade, Alexander Koch, Markus Raiber, Qi Zhao

Studiengang

Informatik

Abschluß

Master

Veranstaltungsart

Seminar

Zyklus

Block

Zusammenfassung

In the setting of secure multiparty computation, two or more parties with private inputs wish to compute some joint
function of their inputs. The security requirements of such a computation are privacy (meaning that the parties learn the
output and nothing more), correctness (meaning that the output is correctly distributed), independence of inputs, and more.
Due to its generality, secure computation is a central tool in cryptography.
In this seminar, we examine modern protocols for secure multiparty computation of arbitrary functions.

Allgemein

Sprache

Englisch

Verfügbarkeit

Zugriff

Unbegrenzt – wenn online geschaltet

Aufnahmeverfahren

Sie können diesem Kurs direkt beitreten.

Zeitraum für Beitritte

Bis: 18. Okt 2021, 08:00

Freie Plätze

Veranstaltungszeitraum

1. Okt 2021 - 31. Mär 2022

Für Kursadministratoren freigegebene Daten

Daten des Persönlichen Profils

Anmeldename
Vorname
Nachname
E-Mail
Matrikelnummer

Zusätzliche Informationen

Objekt-ID

2138993