BRACE is a research-oriented simulation framework for exploring Byzantine-resilient Federated Learning (FL) systems through consensus-based strategies. The project implements and analyzes different configurations of FL systems to evaluate their robustness under adversarial conditions, such as data poisoning and malicious client behavior. The final goal of the project is to define and analyze a consensus-driven protocol to mitigate, and hopefully detect, byzantine attacks.
This project provides a simulation environment where:
- A central aggregator coordinates training across multiple clients.
- A configurable number of Byzantine clients introduce faulty updates to simulate real-world adversarial threats.
- Experiments can be conducted with varying numbers of clients, rounds, selection strategies, and datasets.
The system has been tested on two benchmark datasets:
- MNIST
- FashionMNIST
Two models, one for each dataset, are used and optimized via hyperparameter tuning.
Clone the repository and install the required dependencies:
pip install -r requirements.txtYou can launch the simulation using the federated_learning.py python script.
-t | --task : Task to run (mnist | fashionmnist)
-n | --n_clients : Number of clients (default: 5)
-r | --rounds : Number of rounds (default: 2)
-p | --percentage : Percentage of clients selected per round (default: 1, i.e. all)
-b | --byzantine : Number of Byzantine clients (default: 1)
-rtime | --realtime : Enable real-time training (different local datasets each round)python federated_learning.py -t mnist -n 10 -r 5 -p 0.5
runs the simulation on the MNIST task, with:
-
10 total clients,
-
5 training rounds,
-
50% of clients randomly selected per round.
Two Bash scripts are included to facilitate large-scale experimentation:
-
collect.sh: Automates simulations and collects results for standard FL settings. -
collect_rtime.sh: Automates simulations under real-time training conditions.
Some results have been collected and plotted for visualization into the notebook analysis.ipynb.