TL;DR
This paper proposes a novel hybrid post-quantum cryptography (PQC) approach for securing blockchain-based federated learning (BFL) systems against quantum computer threats. It introduces a multi-factor device role selection mechanism and leverages Verifiable Random Functions (VRFs) for improved consensus, aiming to enhance both security and performance.
What Problem Does This Paper Address?
The advent of quantum computers poses a significant threat to current cryptographic standards like RSA and ECDSA, which are foundational to blockchain and federated learning (FL) systems. This vulnerability necessitates the adoption of post-quantum cryptography (PQC). However, the practical implementation and performance implications of PQC in complex systems like blockchain-based federated learning (BFL) remain largely unexplored. Specifically, the paper addresses two key challenges: 1) designing an efficient and secure PQC framework for BFL against quantum attacks, and 2) tackling performance bottlenecks, particularly in device role selection and consensus mechanisms within BFL.
Methods
The authors propose a hybrid signature approach combining two PQC signature schemes: Dilithium (or Falcon) for its stateless and lattice-based properties, and XMSS (Extended Merkle Signature Scheme) for its stateful, hash-based, and generally smaller signature/public key sizes. Dilithium is used to "certify" the XMSS tree, while XMSS signs the actual transactions. This hybrid strategy aims to leverage the strengths of both while mitigating their individual weaknesses.
For device role selection, a multi-factor dependent fuzzy logic mechanism is introduced. This mechanism calculates a "selection value" for each device based on several factors, including stake, VRF output, contribution value, computational power, shape value (convergence contribution), Wasserstein distance (data distribution similarity), and cosine distance (model similarity). This allows for a more nuanced and efficient assignment of roles (worker, validator, miner) compared to random selection.
To improve consensus mechanisms, the paper proposes integrating Verifiable Random Functions (VRFs). VRFs provide verifiable randomness, ensuring that the selection process is transparent and tamper-proof. This is particularly useful for selecting winning miners in a blockchain setting, aiming to prevent manipulation and improve fairness.
Datasets
The experiments were conducted using the MNIST dataset, a standard benchmark for image classification tasks. The dataset consists of 70,000 samples (28x28 images) with 60,000 for training and 10,000 for testing. The MNIST dataset was sharded across devices to simulate a distributed learning environment.
Key Insights and Ideas
- Hybrid PQC is better than single PQC: Combining Dilithium/Falcon with XMSS offers a more robust security posture and better practical performance than using either scheme alone, addressing issues like large key sizes and state management.
- Multi-factor role selection is crucial: Simply assigning roles randomly in BFL is inefficient. A sophisticated approach considering factors like stake, computational power, and data distribution leads to better system performance and faster learning convergence.
- VRF enhances fairness and efficiency: VRFs can be used to introduce verifiable randomness into critical processes like miner selection, making the consensus mechanism more secure and less prone to attacks.
- PQC impacts performance: The choice of PQC scheme significantly affects communication overhead, block generation time, and overall system accuracy, highlighting the need for careful selection and optimization.
Key Results
- The proposed hybrid signature scheme is demonstrated to be quantum-secure.
- The multi-factor device role selection mechanism showed improved performance and faster learning convergence compared to random selection, particularly in scenarios prioritizing workers.
- Experimental results indicated that the hybrid approach (XMSS + Dilithium/Falcon) offers a good balance between security and performance, with XMSS shining in its small public key size and Dilithium/Falcon offering good key generation and signing performance.
- The paper shows that PQC implementation directly impacts system performance, with some schemes causing significant overhead. The hybrid approach aims to mitigate these issues.
- The proposed VRF-based consensus mechanism is shown to be secure against 51% attacks and reduces the probability of forking events.
Limitations and Future Work
- Assumptions: The current implementation assumes devices do not go offline during a communication round. The authors acknowledge that this approach is tested in a BFL environment and its suitability for other hardware systems or scenarios needs further investigation.
- XMSS Tree Height: The use of smaller XMSS trees (height < 10) is recommended for faster key generation, but this reduces the "life" of a single tree. While new trees can be generated, this trade-off is noted.
- Generalizability: The authors suggest that the proposed role selection mechanism can be further customized to make it more robust and dynamic for different BFL scenarios.
- Further Investigation: The paper notes that the impact of different libraries used for experiments on critical hash storage might require further investigation. The performance of Dilithium and Rainbow schemes in providing overall system performance needs more intensive research.
Summary updated