Blockchain-enabled SplitFed Learning (BSFL)

• The paper introduces BSFL, a framework that removes central servers by employing blockchain-based smart contracts and rotating committees to securely aggregate model updates.
• BSFL leverages sharded split learning to distribute training and communication loads while using median-based consensus for robust model evaluation.
• By integrating performance metrics and security protocols, BSFL enhances resilience against data poisoning and scales effectively with increased shard numbers.

Searching arXiv for the BSFL paper and closely related decentralized Federated Split Learning work to ground the article in current literature. Blockchain-enabled SplitFed Learning (BSFL) is a decentralized distributed learning framework introduced in "Enhancing Split Learning with Sharded and Blockchain-Enabled SplitFed Approaches" (Sokhankhosh et al., 29 Sep 2025). It extends SplitFed Learning (SFL), which combines Split Learning (SL) and Federated Learning (FL), by replacing the centralized server with a blockchain-based architecture that employs a committee-driven consensus mechanism to enhance fairness and security. In the formulation presented in the paper, BSFL builds on Sharded SplitFed Learning (SSFL): SSFL distributes the workload and communication overhead of the SL server across multiple parallel shards, while BSFL further removes the central FL server, introduces smart-contract-governed aggregation and evaluation, and incorporates an evaluation mechanism to exclude poisoned or tampered model updates (Sokhankhosh et al., 29 Sep 2025).

1. Lineage within split and federated learning

BSFL is situated in a sequence of distributed learning designs defined by the trade-offs of FL and SL. FL imposes substantial computational demands on clients, while SL leads to prolonged training times. SFL was introduced as a hybrid approach that combines the strengths of FL and SL: clients train the smaller initial model segment locally, a server trains the heavier segment, and client updates are aggregated by a federated server, for example via FedAvg (Sokhankhosh et al., 29 Sep 2025).

SSFL was proposed to improve SFL scalability by splitting the SL server role across multiple parallel shards, each acting as a mini SFL server, with a federated server further aggregating all shard or server updates. BSFL is the subsequent step in this progression. Its defining move is to replace the centralized server with a blockchain, thereby addressing the remaining centralization in SSFL, including the central FL server as a single point of failure and the associated security and fairness risks (Sokhankhosh et al., 29 Sep 2025).

This progression clarifies a frequent point of confusion: sharding alone does not decentralize SplitFed Learning. In the terminology of the paper, sharding characterizes SSFL, whereas BSFL adds blockchain-based orchestration and committee-driven consensus to obtain end-to-end decentralization.

2. Architecture and protocol mechanics

In BSFL, nodes are divided into shards. Each shard consists of a committee-selected server, described as an SFL server, and multiple clients. Each client trains the client-side neural network segment locally and shares smashed data, namely activations and labels, with its SFL server. The SFL server completes the server-side forward and backward passes and returns feedback gradients to the clients (Sokhankhosh et al., 29 Sep 2025).

The client-side update is written as:

$$W^C_{j,r} \leftarrow W^C_{j,r} - \lambda \cdot \nabla \ell_j(W^C_{j,r})$$

After $R$ rounds, each shard’s server and its clients submit their trained model updates, $W^S_{i,R}$ and $W^C_{j,R}$, to the blockchain ledger via smart contracts. Smart contracts automate model aggregation, update validation and evaluation, consensus on model updates, and reward or incentive distribution (Sokhankhosh et al., 29 Sep 2025).

The committee-driven stage is central to BSFL. Committee nodes, which are the shard servers, download and evaluate model updates using their own local validation-set data. For each submitted pair, the paper gives the validation loss in the form

$$\mathcal{L}_{j,R} = \text{Loss}(Y_{j,R}, \hat{Y}_{j,R})$$

where $\hat{Y}_{j,R}$ is the predicted output. Each committee node assigns a score. For each model, the final score is the median of all committee members’ scores, with the paper explicitly motivating median scoring as robust against outliers and malicious voters. Only the top-$K$ scoring models are selected for aggregation, with the server-side and client-side aggregation rules given as (Sokhankhosh et al., 29 Sep 2025):

$$W^S_{t+1} \leftarrow \frac{1}{K} \sum_{k=1}^{K} W^S_{k,t}$$

$$W^C_{t+1} \leftarrow \frac{1}{K \cdot J} \sum_{k=1}^{K} \sum_{j=1}^{J} W^C_{k,j,t}$$

Committee membership rotates from one training cycle to the next. Nodes are not re-selected in consecutive rounds, and the paper states that new committees are formed based on previous scores. This design links protocol governance to prior performance while attempting to limit collusion and promote fairness (Sokhankhosh et al., 29 Sep 2025).

3. Security, fairness, and scalability properties

The paper presents BSFL as addressing four core challenges inherited by SFL from SL: scalability, performance, fairness, and security. Scalability derives from sharding, which distributes training and communication load across multiple SFL servers and enables parallelism. In the paper’s description, blockchain consensus further decouples and distributes aggregation, with only top updates transmitted and aggregated, limiting communication bottlenecks (Sokhankhosh et al., 29 Sep 2025).

For performance, the design combines sharding with federated averaging layered on top of shards. The paper additionally characterizes the committee evaluation process as committee-based k-fold-like cross-validation, because only high-quality models are included in aggregation. This suggests that BSFL is intended not merely as a coordination substrate but also as a model-selection layer.

Fairness is tied to the absence of a single point of control. Committee rotation, the eventual participation of nodes, median-based evaluation, and transparent smart-contract orchestration are presented as the main fairness mechanisms. The rules for aggregation and committee selection are described as transparent and tamper-resistant because they are executed by smart contracts on the blockchain (Sokhankhosh et al., 29 Sep 2025).

Security is centered on poisoned-update exclusion and resistance to malicious committee behavior. Model updates deemed poisoned, yielding high validation loss or low scoring, are automatically excluded because only the top-$K$ models are aggregated. The paper further states that malicious committee members cannot dominate unless they are the majority, and it specifies the security threshold as a majority of honest committee members, expressed as at least $R$0 honest committee members. Additional conditions given in the paper are at least $R$1 honest clients, if $R$2 is clients per shard and $R$3 is the number of top models kept, and the constraint $R$4 to prevent malicious model aggregation (Sokhankhosh et al., 29 Sep 2025).

A common misconception is that blockchain immutability alone provides model robustness. In the BSFL design, immutability contributes auditability and traceability, but the poisoning defense is specifically attributed to evaluation, median scoring, top-$R$5 selection, committee rotation, and the honest-majority assumption.

4. Experimental evaluation

The paper reports empirical evaluation against baseline SL and SFL approaches on the Fashion MNIST dataset in both 9-node and 36-node setups with non-IID data. It states that SSFL improves performance and scalability by 31.2% and 85.2%, respectively, while BSFL increases resilience to data poisoning attacks by 62.7% and maintains superior performance under normal operating conditions (Sokhankhosh et al., 29 Sep 2025).

The main quantitative results reported in the paper are as follows:

Approach	Normal Test Loss	Attacked Test Loss	Avg. Round Time (min)
SL	0.456	0.981	37.6
SFL	0.430	0.872	37.2
SSFL	0.296	1.010	5.5
BSFL	0.339	0.325	33.7

The reported pattern is structurally important. SSFL achieves the lowest normal test loss and the shortest average round time, which the paper attributes to parallel sharding. BSFL does not match SSFL on round time because blockchain introduces protocol overhead, but it substantially improves attacked test loss relative to SL, SFL, and SSFL. The paper further states that BSFL keeps test loss low when up to 47% of nodes are malicious, whereas SL, SFL, and SSFL degrade sharply under data poisoning (Sokhankhosh et al., 29 Sep 2025).

The paper also states that adding more shards, scaling up to 36 nodes, does not degrade performance or slow convergence for SSFL and BSFL, whereas traditional SL and SFL performance drops at scale. This supports the authors’ claim that sharding and committee-based pruning can preserve learning behavior under larger distributed deployments.

5. Relation to other decentralized SplitFed systems

BSFL belongs to a broader effort to remove central coordinators from federated-split training. A directly relevant comparison is "HLF-FSL. A Decentralized Federated Split Learning Solution for IoT on Hyperledger Fabric" (Penedo et al., 10 Jul 2025), which integrates Federated Split Learning with the permissioned blockchain Hyperledger Fabric. In that system, chaincode orchestrates split model execution and peer-to-peer aggregation without any central coordinator, while transient fields and Private Data Collections are used to keep raw data and model activations private (Penedo et al., 10 Jul 2025).

The architectural contrast is specific. HLF-FSL places orchestration on Hyperledger Fabric, stores only hashes and references on-chain, and relies on peer-to-peer aggregation with endorsement-policy-based confirmation of the global model hash. BSFL, by contrast, is described as replacing the centralized server with a blockchain-based architecture that employs a rotating committee of shard servers to validate, score, and aggregate model updates, with only the top-$R$6 scoring models entering aggregation (Sokhankhosh et al., 29 Sep 2025).

The two systems therefore represent different decentralization strategies within Federated Split Learning. HLF-FSL emphasizes permissioned-blockchain privacy features, including transient fields and Private Data Collections, and reports that it matches centralized FSL accuracy on CIFAR-10 and MNIST while reducing per epoch training time compared to Ethereum-based works (Penedo et al., 10 Jul 2025). BSFL emphasizes committee-driven evaluation, top-$R$7 model filtering, and resilience to data poisoning, and the paper presents it as the first blockchain-enabled framework to implement an end-to-end decentralized SplitFed Learning system (Sokhankhosh et al., 29 Sep 2025).

The paper on BSFL also notes that prior work either focused on federated learning or retained centralized split-learning bottlenecks; hybrid systems such as BlockFeST and work by Sai et al. are mentioned as retaining centralized bottlenecks or failing to address large-scale operation. The BSFL paper positions its novelty in combining sharding, split learning, federated averaging, and blockchain with committee-driven consensus in a single system (Sokhankhosh et al., 29 Sep 2025).

6. Implementation constraints, stopping rules, and research significance

The paper identifies blockchain overhead as an explicit implementation challenge. In the reported implementation, the use of Hyperledger Fabric introduces extra communication and propagation delay, especially as committee size or the number of shards increases. This is reflected in BSFL’s round time, which is higher than SSFL’s even though the paper characterizes it as still competitive (Sokhankhosh et al., 29 Sep 2025).

Committee selection is another practical constraint. The protocol requires an honest majority, and the paper notes future work in making committee member assignment more adaptive to node capacities. Non-IID data heterogeneity is treated as part of the design problem rather than an external nuisance: validating updates across all committee members’ data is described as distributed cross-validation that increases robustness to data heterogeneity (Sokhankhosh et al., 29 Sep 2025).

The stopping criterion also changes in the decentralized setting. For fully decentralized systems like BSFL, early stopping is orchestrated via committee consensus when validation loss worsens. This indicates that decentralization in BSFL extends beyond aggregation to include termination control and model-selection logic (Sokhankhosh et al., 29 Sep 2025).

Within the paper’s framing, BSFL’s significance lies in its attempt to remove the last trusted coordinator from SplitFed-style training while retaining the efficiency rationale of split learning and the stabilizing effect of federated averaging. A plausible implication is that BSFL should be understood not only as a blockchain add-on to SFL, but as a reallocation of control over aggregation, validation, and exclusion decisions from a central server to a rotating, scored, and auditable committee.