Federated learning enables training a global model via clients’ local updates without sharing raw data. However, real-world deployments face critical node-related challenges: on the client side, the lack of reliable incentives may discourage honest participation or even encourage malicious behavior; on the server side, centralized coordination introduces a single point of failure and undermines trust in reward allocation. To address these challenges, we propose GANBFL, an incentive mechanism that integrates GAN-based evaluation with blockchain-assisted coordination. A pre-trained GAN generates a root dataset for contribution evaluation, enabling assessment under Non-IID and privacy-constrained conditions. Client contributions are evaluated based on their performance on root dataset and their reputation. Blockchain smart contracts then coordinate model aggregation and reward allocation according to these evaluations, enhancing the system’s reliability. Experiments under poisoning attack scenarios show that GANBFL facilitates high-quality model aggregation and effectively incentivizes honest participation in federated learning.