Abstract: Federated Learning (FL) allows collaborative model training without sharing raw data, yet it typically relies on a central aggregator that acts as a "black box." This paper introduces MPFL, a protocol leveraging blockchain-based smart contracts and Elliptic Curve Cryptography (ECC) accumulators to provide mathematical "membership proofs." These proofs allow clients to verify their contribution to a global model with minimal overhead—requiring only 96 bytes of storage and 20ms for verification—thereby enhancing transparency, fairness, and auditability in decentralized machine learning.
Introduction
In standard Federated Learning, a central parameter server aggregates model updates from various clients. However, this server is often a trusted third party whose internal operations are opaque. A client has no mathematical guarantee that their specific update was actually included in the final global model. This lack of transparency undermines incentive mechanisms, complicates auditing in regulated industries, and allows for potential bias if a malicious server selectively excludes certain participants.
The MPFL Protocol Architecture
To address this trust gap, researchers from the University of Bologna propose replacing the traditional central server with a blockchain-based architecture. The core of the solution is the MPFL protocol, which utilizes smart contracts to manage the aggregation process and generate cryptographic proofs of inclusion.
As shown in the architecture, multiple clients (Client A, B, ... N) send their local model updates to a smart contract. The contract performs two primary tasks: it aggregates the updates to produce a new global model and simultaneously computes a cryptographic accumulator.
Cryptographic Membership Verification
The technical innovation lies in the use of ECC-based accumulators. Unlike Merkle trees, which grow in size with the number of elements, these accumulators condense a set of model updates into a single, constant-size value. For every update included, the smart contract generates a "witness"—a small piece of data that serves as a proof of membership.
The verification process is efficient and privacy-preserving. As illustrated in the logic flow, a client takes their local model update ($m_a^k$), hashes it ($h()$), and combines it with their received witness ($p_a^k$) using a specific mathematical operation ($*$). If the result matches the global accumulator value ($a^k$), the client has mathematical proof that their contribution was included. Because the protocol operates on hashes, the raw training data remains private.
Experimental Evaluation and Results
The protocol was evaluated using standard benchmarks: Fashion-MNIST (101,770 parameters) and CIFAR-10 using the MobileNetV2 architecture (approx. 2.27 million parameters). Experiments varied the number of clients from 10 to 1,000 to test performance under different levels of complexity.
The results highlight the efficiency of the client-side operations:
- Storage Overhead: Clients only need to store 96 bytes to maintain the necessary verification data, regardless of the total number of participants.
- Verification Latency: The local verification function completes in approximately 20 ms, making it feasible even for resource-constrained IoT devices.
- Scalability: While the smart contract's workload increases with the number of clients (e.g., taking ~1,500 seconds to generate witnesses for 1,000 clients on MobileNetV2), the client-side experience remains constant and extremely fast.
Discussion and Future Work
MPFL provides a robust foundation for verifiable decentralized learning. Beyond simple participation rewards, these proofs are a prerequisite for "Federated Unlearning," where a client might need to prove their data was once part of a model before requesting its removal.
However, challenges remain regarding blockchain throughput. The current implementation on Hyperledger Fabric is suitable for permissioned enterprise environments, but the high gas costs of ECC operations on public blockchains like Ethereum remain a barrier. Future research may explore "universal accumulators" to provide negative proofs (proving an update was not included), further strengthening the auditability of the FL lifecycle.
Conclusion
By integrating cryptographic accumulators with blockchain technology, the MPFL protocol successfully bridges the trust gap in Federated Learning. It provides a scalable, low-overhead solution for membership proof, ensuring that decentralized AI can be both private and transparent.
Summary updated