Description
This talk is about inaccurate assumptions, unrealistic trust models, and flawed methodologies affecting current collaborative machine learning techniques. In the presentation, we cover different security issues concerning both emerging approaches and well-established solutions in privacy-preserving collaborative machine learning. We start by discussing the inherent insecurity of Split Learning and peer-to-peer collaborative learning. Then, we talk about the soundness of current Secure Aggregation protocols in Federated Learning, showing that those do not provide any additional level of privacy to users. Ultimately, the objective of this talk is to highlight the general errors and flawed approaches we all should avoid in devising and implementing "privacy-preserving collaborative machine learning".
Infos pratiques
Prochains exposés
-
[CANCELLED] Black-Box Collision Attacks on Widely Deployed Perceptual Hash Functions and Their Consequences
Orateur : Diane Leblanc-Albarel - KU Leuven
[CANCELLED] Perceptual hash functions identify multimedia content by mapping similar inputs to similar outputs. They are widely used for detecting copyright violations and illegal content but lack transparency, as their design details are typically kept secret. Governments are considering extending the application of these functions to Client-Side Scanning (CSS) for end-to-end encrypted services:[…]-
Cryptography
-
SoSysec
-
Protocols
-
-
A non-comparison oblivious sort and its application to private k-NN
Orateur : Sofiane Azogagh - UQÀM
Sorting is a fundamental subroutine of many algorithms and as such has been studied for decades. A well-known result is the Lower Bound Theorem, which states that no comparison-based sorting algorithm can do better than O(nlog(n)) in the worst case. However, in the fifties, new sorting algorithms that do not rely on comparisons were introduced such as counting sort, which can run in linear time[…]-
Cryptography
-
SoSysec
-
Privacy
-
Databases
-
Secure storage
-