Abstract
In this talk I will present a hybrid peer-to-peer federated learning framework that combines personalisation and reputation-based trust to robustly support context-aware applications such as intrusion detection. The system organises participants in a hybrid overlay that combines a flexible peer-to-peer layer with a tree of superpeers that temporarily coordinate model aggregation. A distributed reputation mechanism evaluates each node’s contributions, promoting reliable participants and limiting the influence of suspicious ones, while personalisation adapts the global model to each client’s local traffic patterns. We evaluate a proof-of-concept implementation on a public network intrusion dataset and show that both personalisation and reputation are important to keep detection accuracy stable when clients have heterogeneous data and some participants behave incorrectly or maliciously.
Bio
Vinicius V. Cogo is an Assistant Professor in the Department of Informatics at the Faculty of Sciences of the University of Lisbon (Ciências ULisboa) and an integrated researcher at LASIGE. He holds a PhD and an MSc in Informatics from Ciências ULisboa and a BSc in Computer Science from the UFSM (Brazil). His research focuses on distributed systems, dependability, fault tolerance, secure storage of critical data, and cloud computing.
