Decentralized Federated Learning with Blockchain

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2156 papers202 conceptsfrozen 2026-10-04

Standard Federated Learning (FL) is built on a paradox: it promises privacy by keeping data on local devices, yet it relies on a central server to manage the entire process. This server acts as a "single point of failure"—if it is compromised, the global model can be poisoned; if it goes offline, training stops. Furthermore, in a centralized system, participants must blindly trust that the server is aggregating updates fairly and not reconstructing their private data.

To solve this, researchers are replacing the central server with a decentralized ledger. By integrating blockchain technology, the role of the "trusted aggregator" is shifted from a single entity to a transparent, immutable protocol.

Replacing the Server with Smart Contracts

In a traditional FL setup, the central server receives model weights, averages them, and sends the new global model back to clients. In a blockchain-enabled framework, this logic is encoded into smart contracts.

As Qin Hu 0001 notes in Blockchained Federated Learning for Internet of Things: A Comprehensive Survey, this architectural shift ensures that the aggregation rules are transparent and immutable. Instead of a server deciding which updates to include, the network reaches a consensus. Dinh C. Nguyen and colleagues describe this in Federated Learning Meets Blockchain in Edge Computing: Opportunities and Challenges as a system where local updates are treated as blockchain transactions. Once verified, these updates are recorded on a shared ledger, providing a tamper-proof audit trail of the model’s evolution.

However, executing complex model averaging directly on a blockchain is computationally expensive. To make this feasible, Leon Witt and Felix Sattler propose using Federated Distillation with highly compressed updates. In Decentralized and Incentivized Federated Learning: A Blockchain-Enabled Framework Utilising Compressed Soft-Labels and Peer Consistency, they demonstrate that by reducing updates to 1-bit compressed soft-labels, the aggregation logic can be executed directly on-chain, eliminating the need for any external coordinator.

Validation through Committee Mechanisms

Without a central authority to vet participants, how do we prevent malicious actors from submitting "garbage" updates that ruin the model? The solution is a decentralized validation layer.

One prominent approach is the Committee Mechanism. In A Decentralized Federated Learning Framework via Committee Mechanism With Convergence Guarantee, Chunjiang Che and Chuan Chen describe a system where a rotating subset of clients acts as a committee. These nodes are responsible for scoring and validating the updates of their peers. By rotating this committee, the system prevents any single node from becoming a permanent bottleneck or a point of corruption.

Nanqing Dong expands on this in Defending Against Poisoning Attacks in Federated Learning With Blockchain by partitioning participants into two roles: proposers (who train) and voters (who evaluate). Voters test the proposed updates against their own local validation data. Only updates that receive a majority of positive votes are aggregated into the global model, creating a robust defense against data poisoning.

Aligning Incentives with Reward and Slash

In centralized FL, participants are often assumed to be altruistic. In the real world, training models is expensive in terms of electricity and compute. Blockchain introduces an economic layer to ensure honesty and participation.

The "Reward-and-Slash" mechanism, proposed in the FLock framework by Nanqing Dong and Jiahao Sun, uses financial stakes to enforce good behavior. Participants must put up a "stake" to join the network. If their updates are voted as malicious or low-quality by the committee, their stake is "slashed" (forfeited). Conversely, high-quality contributors are rewarded with tokens.

This creates a self-organizing ecosystem where being a "good actor" is the most profitable strategy. As Leon Witt highlights in his Systematic Literature Review, this integration of game theory and cryptography allows for "trustless" AI training—where the system works not because we trust the people in it, but because we trust the incentives of the protocol.

From Theory to Swarm Learning

This decentralized paradigm is already seeing success in high-stakes fields like healthcare. Swarm Learning (SL), a specific implementation of blockchain-based decentralized FL, has been used to train models across international borders without moving sensitive patient data.

In Swarm learning for decentralized artificial intelligence in cancer histopathology, Oliver Lester Saldanha and colleagues demonstrate that a decentralized "swarm" of hospitals can achieve the same accuracy as a centralized dataset. By using blockchain to manage the synchronization of model weights, they bypassed the legal and ethical hurdles of data sharing, proving that decentralized orchestration is not just a security upgrade, but a catalyst for collaborative research.

Papers in this version

Showing 60 of 156 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 82 papers
  2. Version 1 · 2026-10-04 · 156 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_decentralized_model_aggregation_ce6b0af2,
  title = {decentralized model aggregation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decentralized-model-aggregation/rdr_ce6b0af2598448509a964874aa023cab},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_decentralized_federated_learning_with_bl_ce6b0af2_v1,
  title = {Decentralized Federated Learning with Blockchain},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decentralized-federated-learning-with-blockchain/rdr_ce6b0af2598448509a964874aa023cab/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}