Research
FedBGS: A Blockchain Approach to Segment Gossip Learning in Decentralized Systems
Standard federated learning keeps data on client devices but still routes all model updates through a central server, which is both a single point of failure and a target for reconstruction…

- arXiv
- 2602.01185
- Published
- 2026-02-01
- Authors
- Fabio Turazza, Marcello Pietri, Marco Picone, Marco Mamei
AI summary
Overview
Research area: Security and privacy in decentralized machine learning — specifically the intersection of federated/gossip learning, blockchain smart contracts, and applied cryptography (arXiv category cs.CR).
Technical level: Advanced. The paper assumes familiarity with federated learning (FedAvg and its variants), gossip learning, Ethereum smart contracts, IPFS content addressing, differential privacy, and partially homomorphic encryption.
Scope: The paper proposes FedBGS, a fully decentralized framework that uses on-chain federated clustering to split clients into homogeneous groups, which then train different slices of a model through asynchronous gossip learning, with blockchain and IPFS handling coordination, storage, and auditability.
What This Paper Is About
Standard federated learning keeps data on client devices but still routes all model updates through a central server, which is both a single point of failure and a target for reconstruction, poisoning, and Sybil attacks. The authors also note that federated learning degrades when client data is non-IID (heterogeneous across clients). FedBGS aims to remove the server entirely by using an Ethereum smart contract as the aggregation and coordination layer, while splitting clients into similar clusters that each train only a portion of the model's final layer, so that heterogeneous data causes less conflict during training.
Key Contributions
- Federated K-Means++ for automatic segmentation. A fully decentralized, one-shot clustering mechanism built on federated analytics that groups clients by their local label distributions without revealing raw data, mitigating non-IID effects.
- Blockchain-based segmented gossip learning. Clients exchange model updates only within dynamically formed segments, with a smart contract defining which range of neurons in the final fully connected layer each cluster is permitted to update.
- Hybrid blockchain and IPFS storage. Aggregation and model storage happen off-chain in IPFS, while references (content identifiers) and the full transaction history are recorded on-chain, addressing the cost and storage limits of public blockchains.
- Empirical validation on non-IID data plus a theoretical comparison. Experiments are reported on CIFAR-10 and MNIST-family datasets, and Table I compares FedBGS against FedAvg, FLchain, BFLC, BlockFL, and BGFL on non-IID handling, hybrid storage, privacy techniques, scalability, and decentralization.
Main Findings
-
FedBGS is presented as the only fully decentralized entry combining all privacy mechanisms. In Table I, the FedBGS row is marked as having robust non-IID handling (one-shot K-Means++ clustering), hybrid Blockchain plus IPFS storage, and HE, DP, and Trimmed-Mean aggregation together, with high decentralized scalability and full decentralization. No other row in the table carries all five.
-
Competing approaches are characterized as partial on at least one axis. FedAvg is described as limited on non-IID data, without hybrid storage, with "basic privacy, vulnerable to inference attacks," high but centralized scalability, and no full decentralization. FLchain is described as moderate on non-IID data, on-chain only, transparency-focused without DP or HE, and limited by a consensus bottleneck. BFLC is described as having limited non-IID handling, on-chain storage with committee consensus, robustness via committee validation, and scalability that is "improved but limited at scale." BlockFL is described as having no special non-IID handling, on-chain only storage with high overhead, transparent validation without DP or encryption, and limited scalability due to consensus overhead. BGFL is described as having limited explicit non-IID handling, on-chain only storage without IPFS, no explicit DP or encryption, but high scalability via gossip communication.
-
The one-shot clustering trade-off is reported as small but is not quantified. The authors state that one-shot methods are used specifically to avoid overloading the blockchain and often compromise clustering accuracy, settling for suboptimal solutions, but that their experiments revealed the resulting performance drop is "generally negligible." No accuracy figures for clustering are given in the supplied text.
-
Numerical results are not reported in the available content. The paper states that convergence is shown graphically through line charts and bar plots, with a table summarizing FedBGS costs on an Ethereum blockchain, but the experimental section is truncated before any accuracy, convergence, or gas-cost values appear. This summary therefore does not report any.
-
Evaluation setup: datasets and models. The paper states that six standard datasets were used, and names MNIST, EMNIST (digits split, approximately 280,000 images total), Fashion-MNIST (60,000 training and 10,000 test images), KMNIST (around 60,000 training and 10,000 test images), and CIFAR-10 (50,000 training and 10,000 test images). Two client models were implemented: NetMNIST (two convolutional layers with 6 and 16 filters of kernel size 5, then fully connected layers of 120 and 84 neurons into 10 classes) and NetCIFAR (3×3 convolutions with 32, 64, and 128 filters, batch normalization, two max-pooling operations reducing feature maps to 8×8, a 256-neuron fully connected layer with dropout, and a 10-class output). All experiments were run locally on an NVIDIA RTX 4080 laptop GPU, and the authors note that STL-10 and ImageNet were not tested due to hardware limitations.
-
Security and scalability are evaluated theoretically, not empirically. The paper explicitly states that evaluations related to security and scalability are conducted at a theoretical level, with decentralization and scalability described as FedBGS's main advantage over the state of the art.
Methodology in Plain English
FedBGS runs in two phases.
Phase 1 — clustering and registration. Every client registers on the blockchain with unique credentials, so unregistered clients cannot participate. Each client computes a summary of the label distribution in its own local dataset and submits it. A one-shot federated K-Means++ procedure groups clients by these summaries. K-Means++ is used because it picks the first centroid at random and then picks subsequent centroids with probability proportional to the squared distance from the nearest existing centroid, which reduces randomness and improves clustering. Because these summaries are small statistical vectors rather than deep model weights, the aggregation can be done inside a smart contract and stored on-chain without excessive fees. The contract then assigns each cluster a neuron index range [s_k, e_k] in the final fully connected layer, and clients query the contract to learn which slice they are allowed to update. Encrypted label distributions are combined with the additively homomorphic Paillier scheme, so the contract never sees raw distributions.
Phase 2 — segmented gossip learning. Clients within a segment train locally and compute gradients, which are clipped and perturbed with Gaussian noise for differential privacy. The noise level is managed by an internal scheduler that decreases linearly over time, reaching its minimum in the final iteration, on the reasoning that early rounds are the most exposed and need the most protection before updates have been mixed with other participants' contributions. Updates are stored on IPFS, with the resulting hash recorded on-chain so any peer can verify integrity by comparing the IPFS hash against the blockchain record; a mismatch triggers a token penalty. At regular intervals the smart contract randomly elects a leader, which collects the privacy-preserving updates from each segment, aggregates them with a trimmed mean (discarding a fixed fraction of the most extreme values to blunt Byzantine contributions), reconstructs the full unsegmented model, stores it on IPFS, and records its content identifier on-chain. Every peer then retrieves the latest model and updates its local copy. Gossip is asynchronous: peers run in independent threads with randomly determined intervals, rather than waiting for synchronized rounds.
Why This Matters
The paper targets a structural weakness rather than a single attack: even privacy-preserving federated learning leaves a central server that can fail or be compromised, and public blockchains are too expensive and too slow to store model parameters directly. FedBGS's contribution is an architecture that keeps the blockchain for what it is good at (identity, coordination, auditability, small statistical aggregates) and pushes bulk data to IPFS, while using clustering to reduce the non-IID degradation that plagues decentralized training.
Potential real-world applications:
- Cross-silo healthcare consortia, where GDPR and HIPAA constrain data sharing and participants need an auditable, non-server record of who contributed what — a setting the paper explicitly frames as its target.
- Multi-organization financial modeling, where no single institution is trusted to hold the aggregation role but a tamper-proof participation and update history is required.
- Edge and IoT deployments, where devices are computationally heterogeneous and an asynchronous gossip scheme with no central aggregator suits variable availability better than synchronized rounds.
- Regulated or auditable machine learning pipelines, where verifiable model provenance (a content identifier on-chain pointing at an IPFS-stored global model) matters independently of accuracy.
Industry relevance: the framework is designed to be compatible with any Ethereum-based blockchain and model-agnostic, so it can be dropped onto existing EVM tooling and IPFS infrastructure. The token-based incentive and penalty system, plus the on-chain client registry, maps onto practical concerns in enterprise deployments: rewarding honest participation, deterring poisoned updates, and preventing Sybil identities.
Future Directions
- Iterative (few-shot) clustering. The authors note that few-shot iterative methods, which fall under the federated learning umbrella rather than federated analytics, remain viable if their blockchain storage trade-offs are carefully assessed. Comparing one-shot against few-shot clustering on accuracy and gas cost is the obvious next experiment.
- Scaling to harder benchmarks. STL-10 and ImageNet were excluded because of hardware limits, so whether the segmentation strategy holds for larger and more complex vision tasks is untested.
- Concrete security and cost measurements. Security and scalability are currently argued theoretically; empirical attack simulations against the DP, HE, and trimmed-mean defenses, plus measured gas costs for registration, clustering, and leader election, would substantiate the Table I claims.
- Tuning the DP scheduler and privacy budget. The linear-decay schedule is motivated qualitatively; how the decay rate interacts with convergence, and how to account for the privacy budget across asynchronous rounds, remains open.
Target Audience
Researchers and graduate students working on decentralized or blockchain-based federated learning, privacy-preserving aggregation, and robust distributed optimization. It is also relevant to practitioners designing cross-silo federated systems for regulated industries who need to understand the concrete engineering trade-offs (gas costs, IPFS offloading, one-shot versus iterative clustering) of replacing a central aggregator with smart contracts. Readers without background in federated learning, differential privacy, or Ethereum will find the paper dense; those seeking quantitative benchmark comparisons will find the supplied content insufficient, since the experimental results are truncated.
Authors’ abstract
Privacy-Preserving Federated Learning (PPFL) is a Decentralized machine learning paradigm that enables multiple participants to collaboratively train a global model without sharing their data with the integration of cryptographic and privacy-based techniques to enhance the security of the global system. This privacy-oriented approach makes PPFL a highly suitable solution for training shared models in sectors where data privacy is a critical concern. In traditional FL, local…
The full abstract on arXiv