Introduction
Federated Learning (FL) has emerged as a cornerstone for privacy-preserving machine learning, allowing multiple clients to collaboratively train a model without sharing their raw data. However, the traditional FL paradigm relies on a central parameter server that acts as a "black box." Clients submit their local updates but receive no cryptographic guarantee that their contributions were actually incorporated into the global model. This lack of transparency leads to "silent failures," potential malicious exclusion of participants, and a lack of auditability in regulated sectors like healthcare.
The MPFL Architecture
To address this trust gap, the Membership Proof in Federated Learning (MPFL) protocol decentralizes the aggregation process. By replacing the central server with a blockchain-based smart contract, the system ensures that the aggregation logic (e.g., FedAvg) is deterministic and publicly verifiable.
As shown in the architecture, clients (A, B, ... N) send their local model updates ($m^t$) to the smart contract. The contract not only computes the new global model ($g^t$) but also maintains a cryptographic accumulator ($a^t$) and generates individual witnesses ($p^t$) for each participant.
Cryptographic Verification
The core innovation of MPFL is the use of Elliptic Curve Cryptography (ECC) based accumulators. Unlike Merkle Trees, where proof size grows logarithmically with the number of participants, ECC accumulators provide constant-size proofs.
The verification process is lightweight and privacy-preserving. A client takes their local update ($m_a^t$), hashes it using SHA-256, and combines it with the witness ($p_a^t$) provided by the smart contract. If the resulting value matches the global accumulator ($a^k$), the client has mathematical proof of their participation. This check is independent of the machine learning model's size, making it suitable for everything from simple linear models to complex deep neural networks.
Performance and Scalability
Empirical evaluations across datasets like Fashion-MNIST and CIFAR-10 (using MobileNetV2 and EfficientNetV2) demonstrate that MPFL is highly efficient:
- Storage: Clients only need to store approximately 96 bytes to maintain their proof.
- Latency: Verification takes roughly 20 milliseconds, enabling near-instant confirmation on edge devices.
- Model Agnostic: The cryptographic overhead remains constant regardless of the number of model parameters, ensuring scalability to large-scale AI applications.
Conclusion
MPFL provides a robust, decentralized solution to the trust problem in Federated Learning. By providing clients with a "receipt" for their work, it paves the way for more transparent incentive systems and verifiable audit trails. Future work may extend this logic to "Federated Unlearning," providing proofs that a client's data has been successfully removed to comply with privacy regulations like GDPR.
Summary updated