Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UnlinkableDFL: a Practical Mixnet Protocol for Churn-Tolerant Decentralized FL Model Sharing

Chao Feng, Thomas Grubl, Jan von der Assen, Sandrin Raphael Hunkeler, Linn Anna Spitz, Gerome Bovet, Burkhard Stiller

TL;DR

This work presents UnlinkableDFL, a DFL framework that combines a peer-based mixnet with fragment-based model aggregation to ensure unlinkability in fully decentralized settings, demonstrating that strong unlinkability can be maintained in decentralized learning workflows.

Abstract

Decentralized Federated Learning (DFL) eliminates the need for a central aggregator, but it can expose communication patterns that reveal participant identities. This work presents UnlinkableDFL, a DFL framework that combines a peer-based mixnet with fragment-based model aggregation to ensure unlinkability in fully decentralized settings. Model updates are divided into encrypted fragments, sent over separate multi-hop paths, and aggregated without using any identity information. A theoretical analysis indicates that relay and end-to-end unlinkability improve with larger mixing sets and longer paths, while convergence remains similar to standard FedAvg. A prototype implementation evaluates learning performance, latency, unlinkability, and resource usage. The results show that UnlinkableDFL converges reliably and adapts to node churn. Communication latency emerges as the main overhead, while memory and CPU usage stay moderate. These findings illustrate the balance between anonymity and system efficiency, demonstrating that strong unlinkability can be maintained in decentralized learning workflows.

UnlinkableDFL: a Practical Mixnet Protocol for Churn-Tolerant Decentralized FL Model Sharing

TL;DR

This work presents UnlinkableDFL, a DFL framework that combines a peer-based mixnet with fragment-based model aggregation to ensure unlinkability in fully decentralized settings, demonstrating that strong unlinkability can be maintained in decentralized learning workflows.

Abstract

Decentralized Federated Learning (DFL) eliminates the need for a central aggregator, but it can expose communication patterns that reveal participant identities. This work presents UnlinkableDFL, a DFL framework that combines a peer-based mixnet with fragment-based model aggregation to ensure unlinkability in fully decentralized settings. Model updates are divided into encrypted fragments, sent over separate multi-hop paths, and aggregated without using any identity information. A theoretical analysis indicates that relay and end-to-end unlinkability improve with larger mixing sets and longer paths, while convergence remains similar to standard FedAvg. A prototype implementation evaluates learning performance, latency, unlinkability, and resource usage. The results show that UnlinkableDFL converges reliably and adapts to node churn. Communication latency emerges as the main overhead, while memory and CPU usage stay moderate. These findings illustrate the balance between anonymity and system efficiency, demonstrating that strong unlinkability can be maintained in decentralized learning workflows.
Paper Structure (46 sections, 5 equations, 9 figures, 1 algorithm)

This paper contains 46 sections, 5 equations, 9 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unlinkability in Federated Learning
  4. Mixnets
  5. Problem Statement and Threat Model
  6. Problem Statement
  7. Unlinkability Game
  8. Adversarial Advantage
  9. Security Notion
  10. Threat Model
  11. Adversary Type
  12. Adversary Capabilities
  13. Adversary Goals
  14. Defensive Objective
  15. UnlinkableDFL
  16. ...and 31 more sections

Figures (9)

  • Figure 1: System architecture of UnlinkableDFL. Each node acts simultaneously as a learner and a mixer. Local models are fragmented, shuffled, and transmitted through a peer-based mixnet with stochastic delays and cover traffic. Neighbor-wise aggregation replaces central coordination, while the management layer handles deployment and monitoring.
  • Figure 2: Local and aggregated accuracy for UnlinkableDFL and Nebula across different numbers of nodes $N$.
  • Figure 3: Aggregated and local accuracy over time in UnlinkableDFL.
  • Figure 4: Relay entropy over the outbox size $O$ (left) and maximum hops $K$ (right).
  • Figure 5: End-to-end path entropy over the outbox size $O$ (left) and maximum hops $K$ (right).
  • ...and 4 more figures