# 26-Spring-BigDataSecurity

**Repository Path**: jqu9/26-spring-big-data-security

## Basic Information

- **Project Name**: 26-Spring-BigDataSecurity
- **Description**: Final Report for Big Data Security course, 2026 Spring, UCAS
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-05-27
- **Last Updated**: 2026-06-02

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# 26 Spring Big Data Security

This repository contains the differential privacy federated learning experiment code for the course project.

## Project Scope

- Dataset: MNIST
- Federated learning model: LeNet
- Differential privacy mechanisms: no DP, Laplace mechanism, Gaussian mechanism
- Privacy accounting: simple composition theorem
- Privacy attack: GRNN gradient inversion attack
- Outputs: training metrics, privacy-utility curves, GRNN reconstruction images, attack metrics, and summary CSV/PNG files

## Main Files

- `大作业.docx`: original course assignment requirement document
- `ASSIGNMENT_REQUIREMENTS.md`: Codex-readable summary of the selected assignment option and report requirements
- `Fed.py`: federated learning training loop with structured metrics output
- `GRNN.py`: GRNN gradient inversion attack with quantitative leakage metrics
- `run_sota_experiments.py`: batch runner for multi-mechanism, multi-epsilon, multi-seed experiments
- `summarize_sota_results.py`: result aggregation and report figure generation
- `CODEX_REMOTE_RUNBOOK.md`: complete remote GPU execution instructions for Codex
- `SOTA_EXPERIMENTS.md`: short command reference

## Remote Execution

Use `CODEX_REMOTE_RUNBOOK.md` as the authoritative runbook for GPU execution. It covers environment setup, smoke tests, full experiment commands, result packaging, and troubleshooting.

## Notes

Datasets, generated results, logs, checkpoints, and archives are intentionally excluded from git. The MNIST dataset is downloaded automatically by the training and attack scripts when needed.