Federated Learning is already a well-established technology widely used by companies such as Google, which employs this approach in training carried out directly on users’ devices. However, in this study, researchers Francinaldo Barbosa (UFPI), Luis Guilherme Silva (UFPI), Iure Fé (UFPI), Israel Cardoso (UFPI), Alex Vieira (UFJF), Geraldo P. Rocha Filho (UESB), and Francisco Airton Silva (UFPI) focused on expanding this architecture by integrating Federated Learning with blockchain, aiming to enhance the security and reliability of the distributed training process.
This paper addresses distributed machine learning, which differs from the traditional machine learning model in that, in the latter, client data is sent to a central server, which has greater computational capacity to process it and train the model. However, this process requires large volumes of data to be transmitted over the network, which can lead to issues related to privacy, security, and the risk of malicious attacks. Federated Learning was developed precisely to minimize these problems. Instead of sending user data to the server, each client trains the model locally using its own data. After this step, only the updated model parameters or weights are sent to the central server, while the data remains stored on the client’s device. This reduces the circulation of sensitive information and preserves user privacy.
To achieve their goal, the authors used FLEX, a framework designed for federated learning simulations. This framework includes various libraries, among them FlexBlock, which is responsible for integrating federated learning with blockchain. This integration allows model updates to be recorded and validated in a decentralized manner, increasing the security of the process and making malicious manipulation more difficult.
The proposed architecture combines two communication models:
Client-server, used during training performed by clients and when sending model updates; Peer-to-peer (P2P), used for communication between the different nodes (servers) in the blockchain network.
In the blockchain architecture, servers participate in a consensus process to decide which update will be incorporated into the next block in the chain. The article compares two consensus mechanisms.
The first mechanism is based on logic similar to Proof of Work (PoW), in which the server that performs the most computational work to validate the block wins.
The second mechanism takes into account the performance achieved during training, prioritizing the server that produced the best results (for example, higher accuracy) when updating the model.
The objective of the study was to compare these mechanisms to identify which one offers better performance and greater potential for application in real-world scenarios of Federated Learning with Blockchain.
Training occurs through successive rounds, a characteristic typical of Federated Learning. In each round, clients train the model locally, send their updated parameters to the server, and a new global model is generated. This process is repeated for a predefined number of rounds.
One of the key metrics analyzed was accuracy, which indicates how much the model improves its performance over the course of training rounds—that is, it measures how much the model has “learned.”
The authors observed that the Proof-of-Work-based mechanism exhibits more pronounced fluctuations in accuracy over time. This occurs because the consensus favors the server that expended the greatest computational effort, and not necessarily the one that produced the best trained model. Thus, it is possible for a server to win the consensus even without having the best training results.
Ideally, the evolution of accuracy is expected to show gradual and stable growth until it reaches a point of stabilization, indicating model convergence.
Another result analyzed is shown in Figure 2 on page 11 (the article is published here), which presents the model’s loss over the rounds. This metric indicates the model’s error during training and allows us to assess its ability to converge: the lower the loss, the better the model’s performance tends to be.
Identified Challenges
The authors highlight some limitations of the study.
The main limitation is that the FLEX framework is still quite new, with little documentation and few published works on its specific libraries, especially FlexBlock. This limitation makes comparisons with other studies difficult and limits the tool’s maturity.
Another point noted is that, although the library provides three consensus mechanisms, only two were used in the research. A third mechanism was not included because it appeared to be outdated or not yet fully functional.
Furthermore, the study could have been enriched by comparing a larger number of consensus mechanisms used in blockchain, thereby broadening the analysis of performance, security, and efficiency.
Future Work
As a continuation of this research, the authors suggest the following steps: evaluate a larger number of performance metrics; compare other consensus mechanisms for blockchain; investigate the computational and energy costs of each mechanism throughout the training rounds; and expand the experiments to scenarios more closely resembling real-world applications of Federated Learning.
Read the full article in the SBRC 2026 proceedings
