Framework for Secure Training Data in Machine Learning
A microservices-based privacy-preserving machine learning framework for secure model training and attack-aware evaluation.
backend
Overview
This framework protects training data in machine-learning workflows by combining privacy-preserving techniques with modular services for training, inference, monitoring, and evaluation. It focuses on practical defenses for sensitive ML systems where model utility, attack resistance, and system observability all matter.
Problem & Goal
Problem:
Machine-learning systems can expose sensitive training data through attacks such as membership inference, gradient leakage, model inversion, and reconstruction. A single privacy technique is often not enough for every deployment scenario, and monolithic implementations make it hard to test, monitor, and extend defenses.
Goal:
The goal was to build an extensible privacy-preserving ML framework that can secure training data, evaluate attack resistance, and support modular experimentation with Federated Learning, Differential Privacy, Homomorphic Encryption, and Secure Multi-Party Computation.
Objectives
- Develop a microservices-based architecture for secure ML training workflows.
- Implement modular Python services for training, inference, attack simulation, performance monitoring, and privacy evaluation.
- Evaluate privacy defenses against membership inference, gradient leakage, model inversion, and related ML attacks.
- Support extensible privacy methods including Federated Learning, Differential Privacy, Homomorphic Encryption, and SMPC.
- Track privacy, utility, and system-performance trade-offs across experiments.
Scope
- Model training and inference services.
- Attack simulation and privacy evaluation workflows.
- Performance monitoring for privacy-preserving machine-learning methods.
- Microservice communication between independently deployable modules.
- Privacy method integration for FL, DP, HE, and SMPC.
- Evaluation reports for attack resistance and model utility.
Tech Stack
The framework is built with Python services and a Flask-based microservices structure. It applies privacy-preserving machine-learning methods, attack simulation modules, and monitoring services to evaluate secure ML workflows across privacy, utility, and performance dimensions.
Screenshots & UI
Key Features
- Privacy-preserving ML workflow for secure training data.
- Modular services for training, inference, monitoring, and privacy evaluation.
- Attack simulation for membership inference, gradient leakage, model inversion, and related threats.
- Support for Federated Learning, Differential Privacy, Homomorphic Encryption, and SMPC.
- Microservices architecture for extensibility and independent experimentation.