Benefiting from the rapidly expanding Internet of Things (IoT) data and powerful computing devices, Artificial Intelligence (AI) can train knowledge and make wise decisions from massive data. In heterogeneous IoT, edge AI devices generate isolated and fragmented federated learning knowledge. To complete more complex tasks, federated learning knowledge sharing must be carried out among edge AI devices. However, traditional ciphertext-policy attribute-based encryption (CP-ABE) sharing technology needs to involve untrusted third parties, which may lead to knowledge deletion, unverifiable access, and single point of failure. In this paper, we propose a knowledge sharing system based on consortium blockchain for secure and verifiable knowledge collaboration. We first design a network-wide consortium blockchain knowledge sharing architecture, in which entities can achieve decentralized and verifiable access control. In addition, to achieve fine-grained access to federated learning knowledge ciphertext while ensuring privacy, we propose a CP-ABE with policy hiding, attribute privacy preservation and revocation, named PHR-CP-ABE, which can ensure the privacy of access policies and attributes, and users whose attributes have been revoked cannot continue to decrypt knowledge. We theoretically verified the security of the federated learning knowledge sharing mechanism. Extensive performance analysis and comparison with related works show that our scheme has lower privacy-preserving computational overhead and communication overhead.