PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Eli Yune, Mahesh Balan

arXiv:2610.17546v1cs.DC

TL;DR

FLAI addresses contribution-accountability and fairness gaps in decentralized federated learning with TEE-backed aggregation, sPEG scoring, and on-chain settlement. It matches FedAvg utility while improving fairness, accountability, and robustness; reported reward fairness is 27% better than reputation-only methods.

  • Problem

    Decentralized federated learning faces unverifiable contribution claims and weak fairness under non-IID data, alongside costly on-chain aggregation and limited accountability.

    Source: §I · p. 1

  • Method

    FLAI combines enclave-validated aggregation, sparse peer evaluation using quality, coverage, promptness, effort, and reputation, and blockchain-based settlement.

    Source: §I · p. 1

  • Results

    FLAI matches FedAvg utility while improving fairness, accountability, and robustness; reward fairness is 27% better (lower Gini coefficient) than reputation-only methods.

    Source: §IX · p. 5, F. Results · p. 5

  • Takeaways & Limitations

    Sealed model serving supports sensitive domains such as healthcare and consumer eHealth, where data locality and individual sovereignty matter.

    Source: A. Motivation · p. 4, §IX · p. 5

    The framework depends on TEE security and attestation integrity, fairness degrades under extreme non-IID class distributions, and scaling beyond thousands of clients requires hierarchical or batched coordination.

    Source: §IX · pp. 5–6

Abstract

from arXiv · show

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 enclave-validated aggregation, sparse Peer Evaluation Graph (sPEG) scoring, and on-chain settlement. sPEG incorporates peer-based quality, class coverage, promptness, effort, and reputation into composite contribution scores. The protocol formalizes contribution evaluation and tokenized incentive distribution, ensuring verifiable fairness and resilience against free-riding and collusion. We evaluate FLAI on image, text, and healthcare datasets and demonstrate parity with FedAvg in utility while improving fairness and accountability. A case study in consumer eHealth (aggregating wearable and electronic health record data for real-world evidence generation) illustrates the protocol's applicability to privacy-critical domains where equitable rewards, auditability, and data sovereignty are essential.

I. INTRODUCTION

The introduction presents FLAI as a TEE-backed decentralized federated learning design combining enclave-validated aggregation, sPEG scoring, and blockchain settlement. It motivates this approach with gaps in accountability and contribution evaluation under heterogeneous data.

A. Federated learning

  • Federated learning shares model updates rather than raw data, but decentralized systems still face unverifiable contribution claims, free-riding and poisoning,, and limited accountability.Blockchain-backed DeFL adds immutable logging and incentives,, while on-chain aggregation costs and fairness under non-IID data remains under-addressed.Source: §I · p. 1

I. INTRODUCTION

  • FLAI combines TEE-backed aggregation, multi-factor sPEG evaluation, formal protocol state machines, and blockchain settlement for decentralized federated learning.The stated evaluation spans image and healthcare datasets and includes a consumer-health real-world evidence case study.Source: §I · p. 1

B. Contribution evaluation

  • Contribution-evaluation methods trade off axiomatic fairness against scale, can be brittle under heterogeneity, or omit class coverage, timeliness, and collusion or Sybil risks, –.The passage identifies Shapley values as poorly scaling,, and influence-function and gradient-similarity methods as potentially brittle,.Source: B. Contribution evaluation · p. 1

DOI: 10.1109/ICCE67443.2026.11449597

The section frames blockchain and TEE roles in decentralized federated learning, then outlines a workflow that assigns coordination and settlement to the chain and scoring and aggregation to enclaves.

C. Blockchain-enabled FL

  • Blockchain-enabled FL offers auditability and programmable payouts,,, but heavy on-chain computation raises latency and gas costs, and rich metadata logging can leak participation patterns.Token-only incentives are described as gameable.Source: C. Blockchain-enabled FL · p. 2

D. TEEs

  • TEEs provide confidentiality, integrity, and attestation,, avoid HE/MPC overhead, but remain exposed to microarchitectural side channels,.FLAI confines scoring and aggregation to enclaves and uses the chain for verification and settlement.Source: D. TEEs · p. 2

A. Entities and workflow

  • Clients submit encrypted local updates for sparse-graph peer evaluation; an enclave verifies reports, scores and aggregates updates, seals the model, and attests, while the blockchain registers tasks, verifies attestations, logs, and settles payouts.The round sequence is local training, peer evaluation, enclave scoring and aggregation, sealed serving, then on-chain attestation and settlement.Source: A. Entities and workflow · p. 2

B. Threats and objectives

The section states FLAI’s threat objectives and describes enclave validation alongside sPEG-based contribution scoring and anti-gaming measures.

B. Threats and objectives

  • The threat model includes poisoning, collusion, Sybil inflation, free-riding, and timing manipulation, with confidentiality, integrity, fairness, and auditability as objectives.The objectives cover data and update confidentiality, scoring and aggregation integrity, fairness across heterogeneous contributors, and immutable logs.Source: B. Threats and objectives · p. 2

A. sPEG Contribution Scoring

  • The score formula multiplies Qi, Ci, (1 + βPi), Ni, and Ri, with a normalized score also specified.The enclave validation module checks range and cross-consistency; scoring and aggregation outputs are sealed and bound to remote attestation, while the host OS and hypervisor are untrusted.Source: A. sPEG Contribution Scoring · p. 2, C. TEE Design and Attestation Flow · p. 2
  • Degree caps and rotation limit mutual score inflation, enclave cross-checks enforce report consistency, and reputation decay reduces stale influence.Source: A. sPEG Contribution Scoring · p. 2

B. Protocol State Machines

The protocol state machines assign separate action sequences to clients, validators, the enclave, and the blockchain.

B. Protocol State Machines

C. Per-Round Workflow

Each round moves from task registration and local training through peer evaluation, enclave scoring and aggregation, attestation, and on-chain settlement. The incentive and security mechanisms combine multidimensional scoring with defenses against manipulation, while retaining explicit privacy and deployment limitations.

C. Per-Round Workflow

  • The round proceeds from task registration and local training to peer evaluation, enclave verification, scoring and aggregation, then attestation and payout settlement.Contracts verify signatures and measurements before settlement; inconsistent or statistically aberrant evaluation reports are discarded.Source: C. Per-Round Workflow · p. 2, §V · p. 3

V. INCENTIVE MECHANISMS AND SECURITY GUARANTEES

  • Rewards use each contributor’s normalized score, while invalid inputs forfeit payment and correlated boosts can trigger slashing.The contracts maintain task, contribution, and settlement records; staking or resource proofs deter Sybils.Source: §V · p. 3
  • Encrypted inputs, sealed storage, and inference-only serving limit leakage, but side-channel risk remains and extreme non-IID or Internet-scale settings pose further constraints.The text identifies hardware trust and side channels as residual risks, disjoint-class distributions as a stress case, and hierarchical enclaves with batched settlements as an Internet-scale need.Source: §V · p. 3
  • Coverage rewards label diversity, promptness rewards timely participation, logarithmic effort limits large-dataset dominance, and reputation rewards sustained quality.Together, these factors are intended to reduce free-riding and late or narrow contributions.Source: §V · p. 3
  • Non-colluding neighbors downscore poisoning updates, while degree caps, rotation, enclave correlation checks, and sparse connectivity constrain collusion and Sybil influence.Weighted aggregation further shrinks the effect of malicious updates.Source: §V · p. 3

VI. IMPLEMENTATION AND OBTAINED RESULTS

The implementation combines SGX enclaves, an Ethereum-compatible testnet, and a sparse peer-evaluation workflow. Reported results preserve benchmark utility while improving fairness and robustness, with measured attack resistance and runtime overhead.

A. Setup and Stack

  • The prototype used Intel SGX with IAS/DCAP attestation, Solidity contracts on an Ethereum-compatible testnet, and gRPC with enclave-anchored TLS.Experiments used 10 SGX servers, 5 s block intervals, up to 500 simulated clients, and CIFAR-10, FEMNIST, and de-identified healthcare data; baselines included FedAvg, FedAvg+Shapley, and blockchain-only incentives.Source: A. Setup and Stack · p. 3

B. sPEG Construction and Aggregation

  • Client profiles combine gradient norms, update magnitudes, label histograms, and local validation metrics; cosine similarity selects k=5 neighbors with capped degrees, refreshed every M rounds.The enclave normalizes peer evaluations, computes scores, aggregates and seals the model, and records measurements on-chain for auditability.Source: B. sPEG Construction and Aggregation · p. 3

C. Smart Contracts and Deployment

  • TaskRegistry, Settlement, and AuditLog are the core contracts, and gas consumption peaks at signature verification and payout loops.Average consumption was ∼210k gas per round.Source: C. Smart Contracts and Deployment · p. 4

D. Results and Analysis

  • 73.1% on CIFAR-10 versus 71.4% FedAvg; FEMNIST was within 0.5%, while fairness and robustness improved.Under 20% label-flipping attacks, attacker scores dropped by 65% and accuracy declined 1.4%; free-riders earned < 1% of rewards and ten-node collusion gained ≤7%. Scoring added 11–18 ms per client, aggregation took < 5 ms for 500 clients, and inference latency averaged 42 ms.Source: D. Results and Analysis · p. 4

E. Ablation Summary

The ablation results quantify how selected scoring components affect reward distribution, participation timing, and free-rider rewards. The application context links decentralized learning to consumer health data streams and local model training.

E. Ablation Summary

  • Removing coverage raises reward Gini from 0.21 to 0.36, removing promptness increases round duration by 17%, and omitting reputation lets free-riders recover 8–10% of total rewards.Source: E. Ablation Summary · p. 4

A. Motivation

  • Centralized RWE collection of EHRs, insurance claims, and observational data raises privacy, consent, and data-ownership concerns; DeFL offers collaboration while preserving data locality.For consumer health devices, the passage describes models trained in users’ sensing and actuation environments.Source: A. Motivation · p. 4

B. Application Context

  • The application combines wearable measurements, personal health records, and optional genomic data, with clients locally training models for adherence, disease-risk, and behavioral predictions.The local computation is described as taking place within secure enclaves on user devices.Source: B. Application Context · p. 4

C. Protocol Deployment

The consumer-health deployment follows the FLAI lifecycle from sponsor tasks and local training through peer evaluation, enclave processing, and sponsor inference. Its rationale emphasizes affordability and inclusion, while the prototype reports predictive, fairness, and latency results.

C. Protocol Deployment

  • Sponsors issue outcome-prediction or behavioral-modeling tasks; devices train locally on multimodal data, validators assess contribution quality, and enclaves aggregate, score, and seal the model.Sponsor inference queries use a pay-per-inference API while raw model parameters and personal data remain inaccessible.Source: C. Protocol Deployment · pp. 4–5

D. Economic and Functional Rationale

  • The proposed economic model shifts computation to user-owned endpoints and charges sponsors for actionable model outputs rather than bulk data procurement and centralized training.The passage presents this as a way to make advanced analytics more accessible to smaller research groups and digital health startups, while scoring is intended to support fair compensation across device types and data volumes.Source: D. Economic and Functional Rationale · p. 5

E. Compliance and Ethics

  • Data remain encrypted on user devices, computations occur in attested TEEs, and on-chain logs track training and reward flows; participants can join or exit through dynamic consent.Source: E. Compliance and Ethics · p. 5

F. Results

  • Predictive accuracy was within 2% of centralized baselines, reward fairness improved by 27% relative to reputation-only methods, and average inference latency was 45 ms per request.The prototype also reported stable embodied learning cycles across heterogeneous devices without data leakage and concluded that embodied federated learning for decentralized RWE generation was feasible.Source: F. Results · p. 5

G. Summary

The case study positions FLAI as a path from privacy-preserving real-world evidence collection to embodied health AI. It also contrasts enclave-based performance with cryptographic approaches and notes a possible hybrid trust design.

G. Summary

  • FLAI is presented as extending privacy-preserving real-world evidence collection to health devices that learn and contribute to global models without exposing personal data.The case study connects on-device intelligence, equitable incentives, and verifiable compliance to health-related AI and real-world evidence.Source: G. Summary · p. 5

A. TEE versus cryptographic approaches

  • TEEs support real-time aggregation and inference compared with HE/MPC, but depend on hardware trust and remain exposed to side channels.A hybrid design could use enclaves for performance-critical paths and ZK/MPC to protect secrets or validate scoring without revealing inputs.Source: A. TEE versus cryptographic approaches · p. 5

B. Marketplace dynamics and generality

FLAI links contribution-based payouts and sealed model serving to marketplace use, while identifying compliance practices, deployment limits, and future design directions.

B. Marketplace dynamics and generality

  • sPEG ties payouts to multidimensional contribution, while sealed serving monetizes model utility via queries.The pattern is proposed for finance, supply chains, and IoT; governance of α, β, and k sets fairness–latency trade-offs and resilience to gaming.Source: B. Marketplace dynamics and generality · p. 5

C. Policy and compliance

  • Attestation formats, consent receipts, and fairness audits are proposed as standards that would enable regulators to verify privacy and equitable reward allocation.The passage states that attested computation and on-chain logs support GDPR/HIPAA alignment.Source: C. Policy and compliance · p. 5

IX. CONCLUSION AND FUTURE DIRECTIONS

  • FLAI matches FedAvg utility while improving fairness, accountability, and robustness, and supports sealed model serving for sensitive domains.Healthcare and consumer eHealth are named as suitable domains.Source: §IX · p. 5
  • The framework depends on TEE security and attestation integrity, while sPEG initialization and topology tuning face cold-start challenges.Fairness degrades under extreme non-IID conditions with disjoint class distributions.Source: §IX · pp. 5–6
  • Scalability beyond thousands of clients requires hierarchical or batched coordination, and governance risks capture without balanced token-weighted and merit-based voting..Source: §IX · p. 6
  • Future directions include hybrid TEE+ZK/MPC trust models, adaptive sPEG topologies, hierarchical enclave rollups, cross-chain settlement, and formal fairness audits.The passage also proposes richer contribution signals and standardized attestations.Source: §IX · p. 6

Related papers

Loading 2610.17546v1…