Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fantastyc: Blockchain-based Federated Learning Made Secure and Practical

William Boitier, Antonella Del Pozzo, Álvaro García-Pérez, Stephane Gazut, Pierre Jobic, Alexis Lemaire, Erwan Mahe, Aurelien Mayoue, Maxence Perion, Tuanir Franca Rezende, Deepika Singh, Sara Tucci-Piergiovanni

TL;DR

This work tackles secure, scalable decentralized Federated Learning by removing the central orchestrator through a blockchain-based framework. It introduces Fantastyc, which offloads validation to a set of servers and anchors off-chain computations with Proof of Availability & Integrity (PoA&I) while storing only cryptographic fingerprints on-chain, coordinated via fault-tolerant distributed storage. The approach achieves enhanced Byzantine tolerance (requiring a majority of honest servers, $2f_s+1$) and maintains practicality by decoupling ordering from data integrity/availability and by using lightweight confidentiality (InstaHide). Experimental results in geo-distributed deployments demonstrate feasible round latency, robust privacy-utility trade-offs, and scalability to large participant pools and sizable models, highlighting Fantastyc’s potential for real-world BC-based FL deployment.

Abstract

Federated Learning is a decentralized framework that enables multiple clients to collaboratively train a machine learning model under the orchestration of a central server without sharing their local data. The centrality of this framework represents a point of failure which is addressed in literature by blockchain-based federated learning approaches. While ensuring a fully-decentralized solution with traceability, such approaches still face several challenges about integrity, confidentiality and scalability to be practically deployed. In this paper, we propose Fantastyc, a solution designed to address these challenges that have been never met together in the state of the art.

Fantastyc: Blockchain-based Federated Learning Made Secure and Practical

TL;DR

This work tackles secure, scalable decentralized Federated Learning by removing the central orchestrator through a blockchain-based framework. It introduces Fantastyc, which offloads validation to a set of servers and anchors off-chain computations with Proof of Availability & Integrity (PoA&I) while storing only cryptographic fingerprints on-chain, coordinated via fault-tolerant distributed storage. The approach achieves enhanced Byzantine tolerance (requiring a majority of honest servers, ) and maintains practicality by decoupling ordering from data integrity/availability and by using lightweight confidentiality (InstaHide). Experimental results in geo-distributed deployments demonstrate feasible round latency, robust privacy-utility trade-offs, and scalability to large participant pools and sizable models, highlighting Fantastyc’s potential for real-world BC-based FL deployment.

Abstract

Federated Learning is a decentralized framework that enables multiple clients to collaboratively train a machine learning model under the orchestration of a central server without sharing their local data. The centrality of this framework represents a point of failure which is addressed in literature by blockchain-based federated learning approaches. While ensuring a fully-decentralized solution with traceability, such approaches still face several challenges about integrity, confidentiality and scalability to be practically deployed. In this paper, we propose Fantastyc, a solution designed to address these challenges that have been never met together in the state of the art.
Paper Structure (29 sections, 4 theorems, 3 equations, 5 figures, 2 tables, 2 algorithms)

This paper contains 29 sections, 4 theorems, 3 equations, 5 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background and Preliminaries
  3. Federated Learning
  4. Blockchain
  5. State of the Art
  6. Design Choices and Key Ideas
  7. Robustness and Enhanced Byzantine Tolerance
  8. Byzantine Clients
  9. Byzantine Servers
  10. Efficient Resource Use by Proof of Availability $\&$ Integrity and Optimally Replicated Fault-tolerant Key-value Store
  11. One blockchain transaction per round
  12. Optimal replication
  13. Confidentiality
  14. System Detailed Design
  15. System model
  16. ...and 14 more sections

Key Result

Lemma 1

(availability and integrity) For any $({{\tt TAG}}, v)$, if exists a valid $P({{\tt TAG}},v)$, then ${{\sf locally\_valid}}({{\tt TAG}}, v)$ holds for at least one correct server, and eventually, $v$ is available from the distributed storage.

Figures (5)

  • Figure 1: Difference Federated Learning (a) and our solution (b).
  • Figure 2: Sequence diagram of the Fantastyc workflow.
  • Figure 3: Study about the impact of multi-client parallelism varying $K$ (while setting $B=5$ ; $E=10$). (a) Test set accuracy vs. communication rounds varying the number of participants (b) Number of communication rounds to reach a target accuracy of 95$\%$ depending on the number of participants per round.
  • Figure 4: Evaluation of the impact of the number of nodes with a constant number of 50 clients.
  • Figure 5: Breakdown of latency within a round with 50 clients.

Theorems & Definitions (8)

  • Lemma 1
  • Proof 1: Sketch
  • Theorem 1
  • Proof 2: Sketch
  • Lemma 2
  • Proof 3: Sketch
  • Theorem 2
  • Proof 4: Sketch