Projects •
GLAM
Full Title
Gossip-based Learning under Adversarial ManipulationDescription
GLAM proposed to investigate adversary models and new security defences in the realm of distributed (gossip-based) machine learning training. The project methodology is two-fold: (i) analysing whether existing techniques are vulnerable to poisoning attacks that can highlight possible violations of data privacy guarantees; and (ii) investigating the possibility of devising training architectures with fewer non-standard networking assumptions. The project aims to produce publications and open-source code that can help to push forward the state-of-the-art in gossip-based learning from security and privacy perspectives.