DS_ppm

Distributed privacy-preserving machine learning simulation with Byzantine fault tolerance experiments on the Adult (Census Income) dataset.

Project Overview

This project simulates a distributed learning setting where:

• each node receives a shard of training data,
• each node applies local anonymization,
• honest nodes train local classifiers,
• Byzantine nodes send malicious outputs,
• a central coordinator aggregates node outputs with different robust strategies.

The goal is to evaluate how faulty/adversarial nodes affect predictive performance and how robust aggregation algorithms can mitigate this impact.

Implemented Components

1) Local Anonymization

The anonymization pipeline is cluster-based and PCA-like:

• data is split into local clusters,
• covariance eigenspace is computed per cluster,
• eigenvector structure is randomized,
• transformed data is mapped back to original dimensionality.

Numerical safeguards are included for tiny clusters to prevent NaN/Inf failures.

2) Distributed Byzantine Simulation

For each experiment:

• number of defective nodes: f in {4, 5, 10},
• total nodes follow n = 3f + 1,
• train/test splits: 60/40, 70/30, 80/20.

Byzantine nodes are simulated by adversarial probability inversion (1 - p).

3) Aggregation Methods

The coordinator compares multiple aggregation rules:

• mean
• median
• trimmed_mean
• krum
• multi_krum
• bulyan (practical variant for this node regime)

Metrics

Each experiment logs:

• accuracy
• f1
• precision
• recall
• cohen_kappa
• roc_auc
• mcc
• time_seconds

Running the Simulation

From the project root:

uv run src/main.py

This generates/updates the experiment output CSV.

Output

Main result file:

• data/byzantine_results.csv

Columns:

• defective_nodes
• total_nodes
• dataset_split
• aggregation_method
• accuracy
• f1
• precision
• recall
• cohen_kappa
• roc_auc
• mcc
• time_seconds

Expected experiment grid size:

• 3 defective-node settings × 3 data splits × 6 aggregators = 54 rows.

Repository Structure

pyproject.toml
README.md
data/
	adult.csv
	byzantine_results.csv
src/
	anon.py
	main.py
docs/
	briefing.txt

Notes

• The project is designed for reproducible experimentation and comparison of robust aggregation methods.
• If you change the Byzantine attack model, method rankings may change significantly.
• main.py is the end-to-end execution entry point for all experiments.