$\texttt{strategic-fl-sim}$: An Extensible Package for Simulating Strategic Behavior in Federated Learning

We introduce $\texttt{strategic-fl-sim}$, a lightweight research package for *strategic behavior* in Federated Learning (FL). Most existing FL frameworks target deployment or other simulation settings—such as privacy, benchmarking or heterogeneity—and are not designed for modeling strategic clients. $\texttt{strategic-fl-sim}$ fills this gap by prioritizing easy specification of **client strategies** (transformations of local updates) and **server defenses** (aggregation methods). The design enforces a clean client–server separation: the $\texttt{Client}$ owns data, local optimizers, local training, and strategic actions, while the $\texttt{Server}$ handles aggregation, global updates and records metrics. Out-of-the-box, $\texttt{strategic-fl-sim}$ includes implementations of common manipulations and aggregators, while remaining extensible. It supports single-node multi-GPU execution for efficient simulations in heterogeneous settings. We showcase the package's utility for the FedAvg protocol and three strategic actions—gradient scaling, sign-flipping, and free-riding.