Trustworthy Cross-Organizational Federated Fine-Tuning of Large Language Models via Blockchain Certification and Resilient Object Storage

In the modern world of artificial intelligence, the most powerful tools are large language models, systems capable of understanding and generating human-like text. These models are trained on vast amounts of data, but that data is often locked away in private silos. Hospitals hold sensitive patient records, banks guard financial histories, and governments protect citizen information. Because of strict privacy laws and security concerns, these organizations cannot simply share their data to train a better, shared intelligence. This creates a dilemma: how can institutions collaborate to improve these powerful tools without ever revealing their private information? The answer lies in a method called federated learning. Instead of moving the data to a central computer, the computer model travels to the data. Each organization trains the model on its own local information and sends only the small, updated mathematical adjustments back to a central coordinator. This keeps the raw data safe, but it introduces a new problem: how do you trust the updates you receive? If an organization sends a corrupted or fake update, or if a file is stolen and altered in transit, the entire shared model could be compromised. Furthermore, storing these massive updates is difficult; they are too large to keep on a public ledger, yet they need a permanent, unchangeable record to prove they are genuine.

Researchers have developed a new framework to solve this trust and storage puzzle, combining two distinct technologies to create a secure system for cross-organizational collaboration. The team, led by Xiaoli Li and colleagues at the Hubei University of Arts and Science, designed a system that splits the job into two parts: a trusted certificate and a secure storage vault. They use a consortium blockchain, which is a shared digital ledger maintained by a group of known, trusted organizations, to act as the certificate. However, instead of writing the massive model files directly onto this ledger—which would be slow, expensive, and technically impossible for large data—they write only a tiny, unique fingerprint of the file. This fingerprint, known as a cryptographic hash, serves as a seal of authenticity. If even a single bit of the model file changes, the fingerprint changes completely, alerting everyone that the file has been tampered with. The actual model data, which can be several megabytes in size, is stored off the blockchain in a resilient, distributed storage system. This separation allows the system to keep the ledger small and fast while ensuring the heavy data remains accessible and secure.

To make this off-chain storage robust, the researchers broke every model file into smaller pieces, or shards, before sending them to storage. Imagine a large document being cut into several pages; if one page is lost or damaged, the others can still be used to recover the whole. These shards are encrypted so that only authorized parties can read them, and they are copied to multiple independent locations. The system is designed so that if one storage node fails, or even if an entire group of nodes goes down due to a power outage or network issue, the other copies in different locations ensure the file is never lost. When an organization wants to retrieve a model update, the system checks the blockchain first to confirm the file's identity and then fetches the shards from the storage network. The recipient can then verify that the pieces match the original fingerprint recorded on the blockchain. This process includes both offline checks, where the file is verified without needing an internet connection, and online checks, which confirm the file is still valid and has not been revoked.

The team tested this framework using a specific type of efficient model update called LoRA, which adjusts only a small portion of a large language model's parameters. They ran experiments with ten participating organizations, simulating a real-world scenario where data is unevenly distributed. They tested different levels of model adjustment to see which produced the best results. The experiments showed that a specific configuration, where the model was adjusted with a rank of eight, achieved the highest performance score on unseen tasks, reaching a score of 4.40 on a ten-point scale. This configuration outperformed other settings, suggesting it offered the best balance for this type of collaborative training. More importantly, the study measured the cost and speed of their new storage method against traditional approaches. They found that storing only the tiny metadata on the blockchain reduced the computational cost, known as gas consumption, by a factor of over 10,000 compared to trying to store the full 8.8 megabyte file directly on the chain. While the full file was stored off-chain, the system also managed to speed up the storage process by nearly 30 percent by uploading the file shards in parallel rather than one by one.

The results demonstrate that separating the proof of a file's identity from the file itself is a practical solution for the future of collaborative artificial intelligence. By keeping the heavy data off the blockchain and using a secure, sharded storage system for the actual files, the researchers created a workflow that is both trustworthy and efficient. The system ensures that every model update can be traced back to its source and verified for integrity without clogging the network or exposing private data. While the total time to complete a full cycle of storage and verification was slightly longer due to the extra steps of sharding and checking, the massive reduction in cost and the increase in reliability make the approach viable for real-world use. The work suggests that as organizations become more willing to collaborate on artificial intelligence, they can do so without sacrificing security or privacy, provided they have a system that can verify the truth of the data without needing to hold all of it in one place. This framework offers a path forward where institutions can build better models together, confident that the contributions are genuine and the data remains safe.