[Paper Review] Federated TrustChain: Blockchain-Enhanced LLM Training and Unlearning

Xuhan Zuo, Minghao Wang+5 more|arXiv (Cornell University)|Jun 6, 2024

Cloud Data Security Solutions4 citations

TL;DR

This paper proposes Federated TrustChain, a blockchain-integrated federated learning framework for Large Language Models (LLMs) that enables transparent, accountable, and efficient unlearning. By leveraging Hyperledger Fabric for tamper-proof logging and Low-Rank Adaptation (LoRA) with tuned hyperparameters, the framework achieves unlearning performance comparable to retraining from scratch—reducing accuracy by only 0.05% on IMDB and 0.19% on Twitter—while maintaining low computational overhead.

ABSTRACT

The development of Large Language Models (LLMs) faces a significant challenge: the exhausting of publicly available fresh data. This is because training a LLM needs a large demanding of new data. Federated learning emerges as a promising solution, enabling collaborative model to contribute their private data to LLM global model. However, integrating federated learning with LLMs introduces new challenges, including the lack of transparency and the need for effective unlearning mechanisms. Transparency is essential to ensuring trust and fairness among participants, while accountability is crucial for deterring malicious behaviour and enabling corrective actions when necessary. To address these challenges, we propose a novel blockchain-based federated learning framework for LLMs that enhances transparency, accountability, and unlearning capabilities. Our framework leverages blockchain technology to create a tamper-proof record of each model's contributions and introduces an innovative unlearning function that seamlessly integrates with the federated learning mechanism. We investigate the impact of Low-Rank Adaptation (LoRA) hyperparameters on unlearning performance and integrate Hyperledger Fabric to ensure the security, transparency, and verifiability of the unlearning process. Through comprehensive experiments and analysis, we showcase the effectiveness of our proposed framework in achieving highly effective unlearning in LLMs trained using federated learning. Our findings highlight the feasibility of integrating blockchain technology into federated learning frameworks for LLMs.

Motivation & Objective

• To address the lack of transparency and accountability in federated LLM training, where private data contributions are hard to verify.
• To enable effective and verifiable unlearning of specific data points from LLMs without retraining from scratch.
• To integrate blockchain technology into federated learning for LLMs to ensure immutability, traceability, and auditability of model updates and unlearning actions.
• To evaluate the impact of LoRA hyperparameters (rank, scaling factor, dropout) on unlearning performance in a blockchain-audited environment.
• To demonstrate the feasibility and efficiency of blockchain-augmented federated learning for LLMs with minimal performance overhead.

Proposed method

• The framework uses a blockchain-based architecture built on Hyperledger Fabric to log all model updates, aggregation steps, and unlearning operations in a tamper-proof, distributed ledger.
• It integrates Low-Rank Adaptation (LoRA) as the fine-tuning mechanism, allowing efficient parameter updates during unlearning by modifying only low-rank matrices.
• Unlearning is implemented by selectively removing or updating LoRA adapters associated with specific data contributions, preserving global model performance.
• The system enforces consensus through endorsement and validation processes in Hyperledger Fabric, ensuring all unlearning actions are cryptographically verified and auditable.
• Hyperparameter tuning of LoRA (rank r, scaling factor alpha, dropout rate) is systematically evaluated to optimize unlearning effectiveness.
• Performance is evaluated using IMDB and Twitter datasets, comparing unlearning accuracy to retraining from scratch and measuring time costs across epochs.

Experimental results

Research questions

• RQ1 How does the integration of blockchain technology improve transparency and accountability in federated LLM training?
• RQ2 To what extent can LoRA-based unlearning achieve performance comparable to retraining from scratch in a federated setting?
• RQ3 What is the impact of LoRA hyperparameters (r, alpha, dropout) on unlearning effectiveness and model accuracy post-unlearning?
• RQ4 What is the computational overhead of integrating blockchain into federated LLM training and unlearning, particularly in terms of setup, transaction processing, and per-epoch time?
• RQ5 Can the blockchain-based system ensure verifiable and auditable unlearning actions without disrupting the training process?

Key findings

• On the IMDB dataset, the best LoRA configuration (r=8, alpha=4, dropout=0.3) achieved a final accuracy of 0.70%, closely matching the 0.65% from retraining from scratch, indicating highly effective unlearning.
• On the Twitter dataset, the best configuration (r=1, alpha=2, dropout=0.3) achieved 8.27% accuracy, compared to 8.08% from retraining, showing minimal performance degradation.
• The average transaction processing time, including model updates and unlearning operations, was 3 seconds, demonstrating efficient handling of LLM-related blockchain transactions.
• Per-epoch training time remained stable at 28–30 seconds, even with unlearning operations, indicating no significant performance degradation.
• The blockchain setup time was 79 seconds at t=0, increasing only slightly to 32.3 seconds per 1000 iterations, showing scalability and manageable overhead.
• The framework successfully created a tamper-proof, verifiable record of all contributions and unlearning actions, enabling full auditability and accountability.