Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust softmax aggregation on blockchain based federated learning with convergence guarantee

Huiyu Wu, Diego Klabjan

TL;DR

This work addresses privacy-preserving distributed learning without a central coordinator by integrating blockchain-based rewards with a robust, softmax-based aggregation in SABFL. Using a PoS blockchain to select validators and miners, SABFL aggregates worker updates through a softmax over approximated population losses, yielding a global model update and incentive-compatible rewards. The authors prove convergence to the global minimum under mild convexity assumptions and demonstrate strong empirical robustness to non-IID data and adversarial behavior, outperforming existing robust aggregation methods across diverse tasks. The approach offers practical scalability on existing blockchains, preserving decentralization while enhancing resilience to data heterogeneity and attacks. Overall, SABFL delivers a theoretically grounded, efficient, and robust BCFL framework suitable for real-world privacy-preserving collaborative learning on decentralized networks.

Abstract

Blockchain based federated learning is a distributed learning scheme that allows model training without participants sharing their local data sets, where the blockchain components eliminate the need for a trusted central server compared to traditional Federated Learning algorithms. In this paper we propose a softmax aggregation blockchain based federated learning framework. First, we propose a new blockchain based federated learning architecture that utilizes the well-tested proof-of-stake consensus mechanism on an existing blockchain network to select validators and miners to aggregate the participants' updates and compute the blocks. Second, to ensure the robustness of the aggregation process, we design a novel softmax aggregation method based on approximated population loss values that relies on our specific blockchain architecture. Additionally, we show our softmax aggregation technique converges to the global minimum in the convex setting with non-restricting assumptions. Our comprehensive experiments show that our framework outperforms existing robust aggregation algorithms in various settings by large margins.

Robust softmax aggregation on blockchain based federated learning with convergence guarantee

TL;DR

This work addresses privacy-preserving distributed learning without a central coordinator by integrating blockchain-based rewards with a robust, softmax-based aggregation in SABFL. Using a PoS blockchain to select validators and miners, SABFL aggregates worker updates through a softmax over approximated population losses, yielding a global model update and incentive-compatible rewards. The authors prove convergence to the global minimum under mild convexity assumptions and demonstrate strong empirical robustness to non-IID data and adversarial behavior, outperforming existing robust aggregation methods across diverse tasks. The approach offers practical scalability on existing blockchains, preserving decentralization while enhancing resilience to data heterogeneity and attacks. Overall, SABFL delivers a theoretically grounded, efficient, and robust BCFL framework suitable for real-world privacy-preserving collaborative learning on decentralized networks.

Abstract

Blockchain based federated learning is a distributed learning scheme that allows model training without participants sharing their local data sets, where the blockchain components eliminate the need for a trusted central server compared to traditional Federated Learning algorithms. In this paper we propose a softmax aggregation blockchain based federated learning framework. First, we propose a new blockchain based federated learning architecture that utilizes the well-tested proof-of-stake consensus mechanism on an existing blockchain network to select validators and miners to aggregate the participants' updates and compute the blocks. Second, to ensure the robustness of the aggregation process, we design a novel softmax aggregation method based on approximated population loss values that relies on our specific blockchain architecture. Additionally, we show our softmax aggregation technique converges to the global minimum in the convex setting with non-restricting assumptions. Our comprehensive experiments show that our framework outperforms existing robust aggregation algorithms in various settings by large margins.
Paper Structure (13 sections, 5 theorems, 19 equations, 3 figures, 4 tables, 1 algorithm)

This paper contains 13 sections, 5 theorems, 19 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. SABFL Framework
  4. Architecture
  5. SABFL
  6. Softmax Aggregation
  7. Analysis
  8. Generalization
  9. Experiments
  10. Design of Experiments
  11. Implementation Details
  12. Results
  13. Conclusion

Key Result

Theorem 1

Suppose Algorithm modified_alg is run with Assumptions validator-var_bound,, decreasing learning rates $\alpha_r$'s, and parameters satisfying for some constant $\delta \in (0,1)$. If $B_r=|\xi_i^{r,t}|$, then the weighted average squared gradient norms satisfy the following bound for all $R \in \mathbb{N}$ Additionally, if then we have

Figures (3)

  • Figure 1: Trends of performance with increasing number of malicious participants on MNIST.
  • Figure 2: Robustness assessment
  • Figure 3: Rankings statistics.

Theorems & Definitions (5)

  • Theorem 1
  • Corollary 1
  • Proposition 1
  • Corollary 2
  • Lemma 1