PoL-BFL: Towards Trustworthy Federated Learning with Zero-Knowledge Proofs and Verifiable Incentives
Decentralized Federated Learning removes the need for a trusted central server but introduces a critical verification gap: participants cannot prove to each other that they have honestly executed the costly local training. This gap undermines both security and incentive alignment, as rational clients are tempted to cheat without fear of automated penalty. Existing defenses are insufficient. Statistical methods cannot cryptographically verify computation, while cryptographic systems incur prohibitive overhead by exhaustively verifying every client. Crucially, both lack a mechanism to automatically convert detection into punishment. To solve the problem, we propose PoL-BFL, a framework that bridges this gap through a novel "verify-then-slash" paradigm. Its core insight is to make verification economically decisive rather than exhaustive. PoL-BFL integrates (1) probabilistic cryptographic verification, where clients efficiently prove training authenticity via zero-knowledge proofs, with only a random subset audited each round; and (2) automated economic enforcement, where smart contracts slash the stake of clients whose proofs fail. This is instantiated in a synergistic three-layer defense that ensures computational authenticity, filters malicious updates that pass cryptographic verification, and enforces incentive compatibility. We prove that honest participation is strictly more profitable than cheating for rational clients whenever the detection probability, stake, and reward parameters satisfy a derived condition. Extensive experiments across eight attack types show that PoL-BFL consistently outperforms single-point defenses. Against free-riding, our detection rate reaches 96.5% compared to 72.5% for ShapleyFL [24]. Against Byzantine attacks, model accuracy under ALIE attack reaches 82.8% versus 68.2% with Krum [5] alone. System overhead remains practical in our controlled testbed: 23.6× faster than Veriblock-FL [9], with per-round gas costs under $1 at 1.5 gwei and $2500/ETH. Code is available at https://github.com/weixiubo/PoL-BFL.