Smart contracts secure over $287 billion in total value locked, yet vulnerabilities caused $3.8 billion in losses during 2023. Traditional detection approaches require complete code disclosure, raising intellectual property concerns for enterprises. We present FedVuln, a privacy-preserving federated graph learning framework enabling collaborative vulnerability detection across mutually distrustful organizations without exposing proprietary code. The framework comprises three specialized components, each targeting distinct challenges in federated smart contract analysis. First, a structured tensor network architecture with unitary constraints achieves 92.3 % centralized accuracy while reducing parameters by 23 % compared to standard tensor GNNs. Second, neural ordinary differential equations, a continuous-time approach justified by hybrid systems theory, capture temporal vulnerability patterns, particularly for reentrancy and MEV attacks, improving detection by 8.3 percentage points for flow-dependent vulnerabilities. Third, divergence-aware aggregation (DAA) operates through geometric interpolation in parameter space, achieving 82.1 % F1-score under extreme non-IID conditions while reducing memory overhead by 34 % compared to SCAFFOLD. Evaluation on 6900 smart contracts across 500 simulated clients demonstrates that FedVuln outperforms established federated baselines by 2.7–4.6 percentage points while reducing communication by 67 %. Privacy analysis confirms 86.7 % accuracy under differential privacy. By enabling privacy-preserving collaboration among blockchain security consortiums, FedVuln provides a practical pathway for organizations to collectively improve smart contract security while protecting intellectual property and maintaining competitive advantages.
