Abstract
Federated Learning (FL) offers a privacy-preserving paradigm for collaborative model training, yet it remains vulnerable to centralized server failures and malicious poisoning attacks. This paper introduces SABFL (Softmax Aggregation on Blockchain-based Federated Learning), a decentralized framework that replaces the central aggregator with a blockchain-based Proof-of-Stake (PoS) mechanism. By employing a novel softmax-based weighting system for model updates, SABFL effectively mitigates the impact of malicious actors and data heterogeneity. Crucially, the framework provides a rigorous mathematical guarantee of convergence under convex conditions. Experimental results across diverse datasets demonstrate that SABFL consistently outperforms existing robust aggregation methods, maintaining high accuracy even in high-adversary environments.
Introduction
Standard Federated Learning (FL) enables multiple participants to train a shared model without exposing raw data. However, its reliance on a central server creates a single point of failure and a target for security breaches. Furthermore, FL is highly susceptible to poisoning attacks, where malicious participants upload corrupted updates to degrade the global model.
While Blockchain-based FL (BCFL) has emerged to decentralize the process, existing solutions often struggle with data heterogeneity—where participants possess varying quantities and distributions of data. Moreover, many BCFL schemes lack theoretical convergence proofs and rely on internal tokens with no real-world value, failing to incentivize honest participation. SABFL addresses these gaps by integrating a robust softmax aggregation mechanism within an established blockchain ecosystem.
Methodology
SABFL decentralizes the aggregation process by utilizing a Proof-of-Stake (PoS) mechanism on a public blockchain (e.g., Ethereum). This ensures that rewards have tangible liquidity. The framework defines three primary roles: Workers (who train), Validators (who verify), and Miners (who aggregate).
Softmax Aggregation
The core innovation is the softmax-based weighting system. Instead of simple averaging, the miner assigns weights to each worker’s update using a softmax function applied to the loss values reported by validators. Validators use their own local data to estimate the "population loss" for each worker's model. If a worker provides a high-quality update (low loss), their weight increases; conversely, poisoned updates (high loss) are mathematically "squashed."
The paper provides a rigorous proof (Theorem 1) demonstrating that this aggregation method converges to the global minimum under convex conditions, even in the presence of noise and varying learning rates.
Experimental Setup
The framework was evaluated across 16 experimental settings, varying the number of participants, the proportion of malicious nodes, and data distribution patterns. The evaluation utilized four diverse datasets and model architectures to ensure generalizability.
The researchers specifically focused on "Label Flipping" attacks, where malicious nodes intentionally mislabel data to subvert the training process. They also investigated the impact of "size heterogeneity," where participants have significantly different sample sizes, which can be as disruptive as intentional attacks.
Results and Discussion
SABFL demonstrated superior performance across nearly all tested scenarios. Excluding "Vanilla" FL (which assumes access to private sample size data that is often unavailable), SABFL was the top performer in 13 out of 16 tasks.
Key Findings:
- Resilience to Attacks: In scenarios where 40% of participants were malicious, SABFL maintained high accuracy, while traditional robust methods like Krum or Median aggregation experienced significant performance degradation.
- Handling Heterogeneity: SABFL proved more effective than standard averaging at handling "size heterogeneity."
- Collusion Resistance: The system remains stable even if up to 50% of validators collude with attackers, as the aggregation relies on a consensus of loss values.
- Consistency: SABFL exhibited the lowest "ranking deviation," indicating reliable performance across different data types and models, unlike other methods that are often dataset-dependent.
Conclusion
SABFL provides a robust and theoretically sound framework for decentralized federated learning. By combining blockchain's security with a novel softmax aggregation mechanism, it effectively counters poisoning attacks and data heterogeneity. While the current convergence proof focuses on convex settings, future work will aim to extend these guarantees to non-convex deep learning models and optimize the computational overhead associated with blockchain transactions.
Summary updated