At the next ROC seminar, we will host a talk by Sara Tucci Head of Laboratory – CEA List, France. His talk will be followed by presentations from two new ROCkers PhD students in the team.
When: January 17, 2025 at 2 pm.
Where: Amphi Friedman, 2 rue Conté, CNAM, Paris, France.
Speaker Sara Tucci, CEA List
Title: Understanding the Ethereum’s Proof-of-Stake Protocol
Abstract: Ethereum’s recent upgrade, known as the Merge, transitioned the blockchain to a Proof-of-Stake model, introducing an original consensus protocol that combines elements of both Nakamoto-style and Byzantine Fault Tolerance designs. This shift led to a complex protocol that, at the time of implementation, was only partially documented. In this talk, I will present our analysis of Ethereum’s Proof-of-Stake protocol, examining its safety, liveness, and incentive compatibility across various network models. Our findings reveal that, in an eventually synchronous network without participant churn, Ethereum’s Proof-of-Stake protocol ensures safety but achieves only probabilistic liveness. Moreover, with the inclusion of the Inactivity Leak mechanism—which removes inactive validators—we identified potential safety vulnerabilities. Finally, we demonstrate that the protocol ensures incentive compatibility in a synchronous setting and achieves eventual incentive compatibility in an eventually synchronous environment.
Speaker: Maham Kayani
Title:Enhancing Anomaly Detection: A Modified Approach to the Dinatrax Framework
Abstract:Dinatrax is a model for detecting anomalies in networks. We’re exploring new approaches within the Dinatrax framework to examine their potential impact and to discover even better results, aiming to further advance the field of network anomaly detection
Bio: I am Maham, a PhD student working in the field of network anomaly detection. Madame Selma is my thesis supervisor.
Speaker: Salah Bin Ruba
Title: Security Issues in Federated Learning for Anomaly Detection in xG
Abstract: Federated Learning (FL) has become a groundbreaking approach to distributed machine learning, enabling participants to collaboratively train models while maintaining data decentralization and privacy. Despite its inherent privacy advantages, FL faces significant security challenges that can threaten the integrity, confidentiality, and availability of the learning process. These challenges include data heterogeneity, communication bottlenecks, and various security threats, such as poisoning attacks, model inversion, and Byzantine behaviors, all of which can compromise model performance or expose sensitive information.
Bio: Salah Bin-Ruba is a PhD candidate at Cedric Lab, CNAM, Paris. Holds an M.Sc. in Data Science and Network intelligence from Telecom SudParis. He is currently working on Federated Learning security in 5G infrastructure.
