FedGuard reproduction notes =========================== Scope: reproduce the centralized PaySim and IEEE-CIS baseline. Datasets are processed independently; this is not federated training. Dataset acquisition 1. Download the PaySim CSV from https://www.kaggle.com/datasets/ealaxi/paysim1. 2. Download train_transaction.csv and train_identity.csv from https://www.kaggle.com/competitions/ieee-fraud-detection/data. Access may require a Kaggle account and acceptance of competition terms. 3. Place files under data/raw/ with the filenames shown in data/README.md. The pipeline uses labeled training files and does not use the Kaggle competition unlabeled test set. Environment and commands The recorded environment was Python 3.11.5 on Windows 10. FedGuard_Training_Environment_Requirements.txt is the captured pip freeze. Install with: python -m pip install -r FedGuard_Training_Environment_Requirements.txt From the repository root run: python scripts/prepare_stage1.py Then run: python scripts/train_stage1.py --dataset paysim --no-wandb Then run: python scripts/train_stage1.py --dataset ieee_cis --no-wandb Preparation makes seed-42 stratified 70/15/15 splits before fitting feature filters and transformations. Training fits/resamples only training data, selects models and thresholds using validation, then evaluates each selected model once on its held-out test split. Reports are written to reports/. Hardware and library differences may cause small numeric variations. Supporting results FedGuard_Supplementary_Validation_Results.csv retains the best validation strategy for each model family and benchmark, including threshold-based validation metrics. It is an eight-row family summary, not a row-by-row record of all 195 candidate fits. FedGuard_Supplementary_Strategy_Validation_Results.csv reports the highest-validation-AP candidate for each model-family and imbalance-strategy search (20 rows across PaySim and IEEE-CIS). These are validation-only comparison summaries, not all 195 individual candidate records; no test metrics are included in these rows. FedGuard_IEEE_CIS_SMOTE_ENN_Validation_Candidates.csv contains all 39 candidates from the separate IEEE-CIS SMOTE-ENN validation-only experiment, with candidate parameters, validation AP and ROC-AUC, and family-winner flags. The test data were not loaded or evaluated for this experiment. FedGuard_Feature_Inventory.csv lists included, excluded, dropped, engineered, indicator, and transformed features. FedGuard_Training_Environment_Requirements.txt is the captured pip freeze for the recorded environment.