Simulations

Reusable simulations of Byzantine-resilient learning protocols.

Two architectures are supported: a single parameter-server that aggregates \(n\) worker gradients each round, and a peer-to-peer topology where each honest worker mixes its model with those received from neighbors. Both share the same primitives (aggregators, attacks, models) but differ in the training loop structure and worker interaction pattern.

Available simulations

• Centralised simulations
• Decentralised simulations