PRISM-FL

This repository contains the experimental code for PRISM-FL, a scalar-verification framework for privacy-aware malicious-update detection in federated learning.

Repository Layout

• main.py: main experiment entry point.
• config/: YAML experiment configurations.
• experiments/: scripts for the paper experiments, ablations, sensitivity studies, and benchmark runs.
• aggregation/: FedAvg, Krum, Median, FLTrust, and PRISM-FL aggregation logic.
• attacks/: gradient flipping, scaling, label flipping, backdoor, noise, and adaptive attacks.
• defense/, privacy/, crypto/: verification, privacy, and simulated secure-protocol components.
• fl_core/: client/server training loop.
• models/: model definitions.
• data/: data loaders and partitioning utilities.
• visualization/: plotting scripts for experimental figures.
• utils/: logging, metrics, I/O, and reproducibility helpers.

Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

On Windows PowerShell:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Quick Start

Run a smoke-style experiment with the default configuration:

python main.py --config config/default.yaml

Run the bundled experiment launcher:

python experiments/run_all.py --config config/default.yaml --exp all

Most paper-specific runs are under experiments/, including run_journal_suite.py, run_v18_atomic.py, and the run_v18_*.sh launch scripts.

Data

Large raw datasets are not included in this repository. Place local VeReMi data under:

data/veremi_extension/

The loaders in data/ can also generate synthetic VeReMi-like data for quick tests where supported by the selected configuration.

Outputs

Experiment outputs are written to results/ by default. This directory is ignored by Git because it can contain large logs, CSV files, and model artifacts.

Reproducibility Notes

• Set random seeds through the configuration files where applicable.
• Keep generated results, model checkpoints, and downloaded datasets out of version control.
• Use the exact configuration file and script name when reporting a reproduced table or figure.