Skip to content

BlindFL: Segmented Federated Learning with Fully Homomorphic Encryption

Authors: Evan Gronberg; Liv d'Aliberti; Magnus Saebo; Aurora Hook
Year: 2025
Venue: arXiv preprint

type: system collection: ppml status: auto

Links: arXiv:2501.11659

Summary

Auto-extracted summary (unreviewed)

BlindFL is a federated learning framework in which each client encrypts with single-key FHE only a chosen subset of its model's parameter matrices and sends them for aggregation, keeping aggregated updates fully encrypted while cutting per-client transmission and compute. On LeNet-5/MNIST and ResNet-20/CIFAR-10 it roughly halves to thirds the round time of plain FHE-based FL with under 1 point accuracy loss, and the segmentation also impedes client-side model poisoning and gradient inversion attacks.

Abstract

Federated learning (FL) is a popular privacy-preserving edge-to-cloud technique used for training and deploying artificial intelligence (AI) models on edge devices. FL aims to secure local client data while also collaboratively training a global model. Under standard FL, clients within the federation send model updates, derived from local data, to a central server for aggregation into a global model. However, extensive research has demonstrated that private data can be reliably reconstructed from these model updates using gradient inversion attacks (GIAs). To protect client data from server-side GIAs, previous FL schemes have employed fully homomorphic encryption (FHE) to secure model updates while still enabling popular aggregation methods. However, current FHE-based FL schemes either incur substantial computational overhead or trade security and/or model accuracy for efficiency. We introduce BlindFL, a framework for global model aggregation in which clients encrypt and send a subset of their local model update. With choice over the subset size, BlindFL offers flexible efficiency gains while preserving full encryption of aggregated updates. Moreover, we demonstrate that implementing BlindFL can substantially lower space and time transmission costs per client, compared with plain FL with FHE, while maintaining global model accuracy. BlindFL also offers additional depth of security. While current single-key, FHE-based FL schemes explicitly defend against server-side adversaries, they do not address the realistic threat of malicious clients within the federation. By contrast, we theoretically and experimentally demonstrate that BlindFL significantly impedes client-side model poisoning attacks, a first for single-key, FHE-based FL schemes.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution segmented FL aggregation where clients send an FHE-encrypted subset of parameter matrices llm 0.85
contribution tunable subset size trading efficiency against full-model participation llm 0.85
contribution reduced per-client space and time transmission costs compared with FL with full FHE llm 0.85
contribution theoretical and experimental defense against client-side model poisoning in single-key FHE FL llm 0.85
contribution evaluation of gradient inversion attack success versus number of shared layers llm 0.85
schemes CKKS llm 0.85
libraries Pyfhel llm 0.85
libraries SEAL llm 0.85
libraries Flower llm 0.85
models LeNet-5 llm 0.85
models ResNet-20 llm 0.85
datasets MNIST llm 0.85
datasets CIFAR-10 llm 0.85
tasks federated learning llm 0.85
tasks image classification llm 0.85
tasks defense against gradient inversion and model poisoning attacks llm 0.85
security level 128-bit llm 0.85
threat model server-side adversary (gradient inversion) and malicious clients (model poisoning); single-key FHE llm 0.85
interaction other llm 0.85
code url https://github.com/ibarrond/Pyfhel (URL found in the full text) regex 0.50
code url https://github.com/adap/flower (URL found in the full text) regex 0.50
code url https://github.com/Microsoft/SEAL.Microsoft (URL found in the full text) regex 0.50
limitations FHE ciphertext expansion increases transmitted data by roughly 27x over plaintext llm 0.85
limitations models with many small parameter matrices (ResNet-20) suffer greater space overhead because each matrix is encrypted individually llm 0.85
llm relevance low llm 0.85
key result MNIST test accuracy / round time: 98.53% / 49.94 s (FHE on, segmentation on) vs 98.82% / 139.18 s (FHE on, segmentation off) (LeNet-5, 3 parameter matrices per client, averaged over 5 runs of 50 rounds) llm 0.85
key result CIFAR-10 test accuracy / round time: 77.53% / 102.31 s (FHE on, segmentation on) vs 78.57% / 275.91 s (FHE on, segmentation off) (ResNet-20, 30% of parameter matrices per client, 5 runs of 50 rounds) llm 0.85
key result data sent per client with FHE: 5,977 KB (1 layer shared) vs 59,777 KB (all 10 layers) (LeNet-5 on MNIST, 10 clients) llm 0.85
key result data sent per client with FHE: 12,348 KB (10% of model) vs 122,701 KB (100%) (ResNet-20 on CIFAR-10) llm 0.85

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: CKKS 6 [p.4], Differential privacy 4 [p.3]
  • library: Pyfhel 11 [p.7], Microsoft SEAL 4 [p.7], PyTorch 1 [p.3]
  • model: LeNet 12 [p.7], ResNet 9 [p.7], GAN 1 [p.13]
  • dataset: MNIST 23 [p.2], CIFAR-10 19 [p.2]
  • technique: Federated learning 31 [p.1], Malicious model 24 [p.1], Secure aggregation 5 [p.1], Semi-honest model 3 [p.3], Pruning 2 [p.2], Fine-tuning 2 [p.1], Model inversion 2 [p.13]
  • assumption: 128-bit security 3 [p.7]
  • task: Medical / healthcare 3 [p.1], IoT / edge 3 [p.1]
  • topic: Homomorphic encryption 101 [p.1], Open source 3 [p.13]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/ppml/federated-learning/gronberg-2025_blindfl-segmented-federated-learning-with-fhe.pdf ppml/federated-learning 13 2.7 MB d9a8a146fecf was Gronberg et al. - 2025 - BlindFL Segmented Federated Learning with Fully Homomorphic Encryption.pdf

Database key: blindfl-segmented-federated-learning-with-fully-homomorphic-encryption · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set blindfl-segmented-federated-learning-with-fully-homomorphic-encryption COLUMN VALUE