Skip to content
Crypto.world

Research

FLAI Protocol: Decentralized Federated Learning with On-Chain Rewards and sPEG-Based Contribution Scoring

Federated learning lets many participants train a shared model without pooling their raw data, but it has no reliable way to tell who actually contributed useful work, so participants can free-ride…

FLAI Protocol: Decentralized Federated Learning with On-Chain Rewards and sPEG-Based Contribution Scoring
arXiv
2610.17546
Published
2026-10-13
Authors
Eli Yune, Mahesh Balan

AI summary

Overview

Research area: Decentralized federated learning, with a focus on incentive design, hardware-rooted trust (trusted execution environments), and blockchain settlement.

Technical level: Advanced. The paper assumes familiarity with federated learning, Intel SGX attestation, smart contracts, and multi-factor scoring schemes.

Scope: The paper specifies, implements, and evaluates the FLAI Protocol — a TEE-backed decentralized federated learning framework that scores client contributions through a sparse Peer Evaluation Graph (sPEG) and distributes tokenized rewards on-chain — including a consumer eHealth case study.

What This Paper Is About

Federated learning lets many participants train a shared model without pooling their raw data, but it has no reliable way to tell who actually contributed useful work, so participants can free-ride, poison updates, or inflate their own claims. Existing blockchain-based decentralized federated learning (DeFL) systems add logging and payouts but suffer from expensive on-chain computation, unverifiable self-reported contributions, and weak fairness guarantees under heterogeneous non-IID data. The FLAI Protocol addresses this by confining scoring and aggregation inside hardware enclaves, scoring clients on multiple peer-evaluated factors, and settling rewards through smart contracts.

Key Contributions

  1. A TEE-backed DeFL design in which an enclave validates inputs, computes contribution scores, aggregates model updates, seals the resulting model, and exposes an inference-only serving API bound to enclave measurements (MRENCLAVE).
  2. sPEG (sparse Peer Evaluation Graph): a neighbor-limited peer evaluation scheme that combines five factors — quality, class coverage, promptness, effort, and reputation — into a composite score that directly weights aggregation and payouts.
  3. A formal specification of the client, validator, enclave, and blockchain state machines, plus a per-round workflow from task posting through attestation and settlement.
  4. A security, privacy, and incentive analysis covering poisoning, collusion, Sybil attacks, free-riding, and timing manipulation, backed by an SGX/Ethereum-testnet prototype evaluated on CIFAR-10, FEMNIST, and de-identified healthcare data, with a full consumer health real-world-evidence (RWE) case study.

Main Findings

  • Utility parity with FedAvg: FLAI reaches 73.1% on CIFAR-10 versus 71.4% for FedAvg, and stays within 0.5% on FEMNIST, while improving fairness and robustness.
  • Attack resistance: Under 20% label-flipping attacks, attacker scores drop by 65% and accuracy declines only 1.4%.
  • Free-riding and collusion are penalized: Free-riders earn <1% of rewards, and a ten-node collusion gains ≤7%.
  • Score factors behave as intended: Promptness yields up to 12% bonuses, and effort correlates with dataset size (r = 0.78); the normalized score Ŝᵢ correlates with marginal utility.
  • Ablations isolate each factor: Removing coverage Cᵢ raises the reward Gini from 0.21 to 0.36; removing promptness Pᵢ increases round duration by 17%; omitting reputation Rᵢ lets free-riders recover 8–10% of total rewards.
  • Practical overheads: Enclave scoring adds 11–18 ms per client, aggregation takes <5 ms for 500 clients, inference latency averages 42 ms, and gas averages ~210k per round.
  • Healthcare case study results: Predictive accuracy within 2% of centralized baselines, reward fairness improved by 27% (lower Gini) relative to reputation-only methods, and secure inference averaging 45 ms per request.

Methodology in Plain English

Clients train models locally and send encrypted updates rather than data. Instead of trusting self-reports, each client builds a profile from signals like gradient norms, update magnitudes, label histograms, and local validation metrics; cosine similarity over these profiles selects a small set of neighbors (k = 5) that evaluate each other, with degree caps and periodic rotation to prevent cliques from forming and mutually inflating scores. Those peer evaluations are combined with class coverage, timeliness, effort (a logarithmic function of update size, so very large datasets do not dominate), and a decaying reputation term into one composite score. The score does double duty: it weights how much each client's update counts in the next global model, and it sets that client's share of the round's reward pool.

All scoring and aggregation happen inside Intel SGX enclaves that verify inputs, cross-check reports for consistency, seal the model, and emit a remote attestation. Solidity contracts on an Ethereum-compatible testnet register tasks, verify attestations and enclave measurements, record scores and model hashes in an audit log, and distribute payouts. The prototype runs on 10 SGX servers (Xeon E-2288G, 64 GB RAM) with 5 s block intervals on a local Geth testnet, simulating up to 500 clients, and is compared against FedAvg, FedAvg+Shapley, and blockchain-only incentive baselines.

Why This Matters

The paper treats fairness in federated learning as a measurable, enforceable property rather than a side effect, and it shows that hardware attestation can replace expensive cryptographic alternatives for the scoring and aggregation path while preserving auditability.

Impact on research: It provides a concrete multi-factor alternative to Shapley-value attribution (which the authors note scales poorly) and couples contribution scoring directly to aggregation weights and payouts, giving future work a benchmark and a formal state-machine specification to build on.

Real-world applications:

  • Consumer eHealth and real-world evidence generation from wearables, personal health records, and optional genomic data, as demonstrated in the case study.
  • Financial fraud detection across institutions without centralizing transaction data.
  • Confidential demand forecasting in supply chains.
  • Distributed anomaly detection across IoT devices.

Industry relevance: The sealed pay-per-inference API turns a trained model into a monetizable service without exposing parameters, which is directly relevant to organizations that cannot legally pool data under GDPR, HIPAA, or PIPA but still want shared analytics — and the stated economic rationale is that shifting computation to user-owned endpoints lowers the fixed cost of bulk data procurement.

Future Directions

  • Hybrid trust models: combining TEEs with zero-knowledge proofs or secure multi-party computation so that performance-critical paths stay in enclaves while secrets or scoring validity are protected cryptographically, reducing reliance on hardware trust alone.
  • Richer contribution signals and adaptive topology: incorporating novelty, entropy, and verification difficulty, and addressing sPEG's cold-start and topology-tuning problems.
  • Scaling and settlement: hierarchical enclave rollups, batched settlements, and cross-chain settlement to move beyond the current prototype scale of hundreds of clients toward thousands.
  • Governance and compliance standards: mitigating governance capture through balanced token-weighted and merit-based voting, and standardizing attestation formats, consent receipts, and formal fairness audits so regulators can verify equitable reward allocation, not just privacy — with integration into the Model Context Protocol and agentic AI workflows as a longer-term direction.

Target Audience

Researchers and engineers working on decentralized or blockchain-based federated learning, incentive mechanism design, and privacy-preserving machine learning will get the most from this paper. It also suits practitioners in regulated health and finance domains who need auditable, hardware-attested collaboration across data silos, and graduate students looking for a worked example of combining TEEs with on-chain settlement. Readers without background in SGX attestation or smart contracts should expect a steep entry curve.

Authors’ abstract

Federated learning (FL) enables collaborative model training without data centralization but struggles with incentive alignment, accountability, and robustness under heterogeneous non-IID data. Existing decentralized FL (DeFL) systems leverage blockchains for coordination yet face high on-chain overhead, unverifiable self-reports, and weak fairness guarantees. We present the FLAI Protocol, a trusted execution environment (TEE)-backed DeFL framework that integrates…

The full abstract on arXiv
Read the original paper