FEDERATED LEARNING & ENCRYPTION

Undergraduate thesis exploring privacy-utility trade-offs in federated learning using homomorphic encryption, differential privacy, and a novel sequential hybrid approach.

PyTorchOpacusPyfhelDockerCIFAR-10

GitHub

Overview

This undergraduate thesis systematically evaluates software-based privacy mechanisms in federated learning. We implemented and compared five FL configurations — baseline, homomorphic encryption (HE), differential privacy (DP), standard hybrid (simultaneous HE+DP), and a novel sequential hybrid approach (DP during training, HE during aggregation) — on CIFAR-10 image classification. The sequential hybrid is our key contribution: by temporally separating privacy mechanisms, it prevents error compounding and improves learning stability over the standard simultaneous application. All experiments were containerized with Docker for reproducibility.

Key Results

68.34%

Baseline Accuracy

57.90%

HE Accuracy

35.05%

DP Accuracy

33.51%

Sequential Hybrid

92.1%

HE non-IID Retention

94%

HE Comm. Reduction

Methodology

Built a modular FL framework in PyTorch with a client-server architecture using FedAvg. Implemented DP via Opacus (gradient clipping + calibrated Gaussian noise) and HE via Pyfhel (CKKS scheme with parameter quantization and chunking). Evaluated across IID and non-IID (Dirichlet α=0.5, α=0.1) data distributions with 10 clients. Privacy was measured through membership inference attacks (threshold-based and shadow model approaches). Resource overhead tracked computation time, communication costs, and memory.

What We Built

●Five FL configurations evaluated under consistent experimental conditions
●Novel sequential hybrid approach — DP during training, HE during aggregation
●Comprehensive MIA-based privacy evaluation (threshold + shadow model)
●IID and non-IID data distribution analysis (Dirichlet α=0.5, α=0.1)
●Full resource overhead analysis: computation, communication, memory
●Dockerized reproducible experimental framework

Challenges

●Balancing computational overhead of CKKS encryption with training efficiency on CIFAR-10 CNNs
●Achieving meaningful privacy guarantees while maintaining model utility across five distinct configurations
●Running concurrent Lambda-style experiments locally to simulate distributed FL clients
●Calibrating differential privacy noise multipliers to find the privacy-utility sweet spot

Outcomes

●Quantified a clear inverse relationship between privacy protection and model accuracy across all five approaches
●Demonstrated the sequential hybrid improves learning stability over standard simultaneous application
●Showed HE maintains 92.1% accuracy under highly non-IID conditions, while hybrid approaches retain only 56–66%
●Published as an IEEE conference paper: 'Privacy-Utility Trade-offs in Federated Learning for 6G Networks'

Papers & Reports

Privacy-Utility Trade-offs in Federated Learning for 6G Networks: A Systematic Evaluation of Software-based Privacy Mechanisms

IEEE conference paper presenting the systematic evaluation framework and sequential hybrid approach for privacy-preserving FL in next-generation networks.

Exploring Federated Learning and Encryption — Undergraduate Thesis

Full 33-page thesis with comprehensive background, methodology, results, and discussion on privacy-preserving federated learning.

Back to all projects