Abstract: Aimed at the key problems of existing cryptographic federated learning schemes, including low system entity trustworthiness, encryption efficiency bottlenecks, single-point-of-failure risks of centralized storage, and insufficient robustness resulting from malicious node poisoning, we present a trusted federated learning scheme based on dynamic searchable hybrid encryption. This scheme focuses on mitigating external untrusted entity threats and high privacy budget conflicts within distributed environments. By incorporating dynamic searchable encryption and hybrid cryptographic mechanisms, it constructs a decentralized, trusted federated learning system featuring high entity trustworthiness, while simultaneously balancing overall system robustness and operational performance. Firstly, we introduce a lightweight bidirectional authentication and distributed key negotiation protocol that enables third-party key management entities and clients to compute negotiated keys with low interaction overhead. This is complemented by a hybrid encryption mechanism to achieve differentiated privacy protection and distributed key management for data. Secondly, by integrating contribution-aware local evaluation with a secure aggregation algorithm based on hybrid encryption, we realize fair and secure ciphertext parameter aggregation. Finally, through security token-driven access control and dynamic key rotation mechanisms, we construct searchable associative trapdoor indexes for centralized ciphertext storage, ensuring non-replayability and non-revealability of historical data. Experimental results show that compared with the existing schemes, the global model accuracy of the MNIST dataset and Fashion-MNIST dataset can reach 99.51% and 93.91% respectively, and the average encryption time overhead of the proposed scheme for private data is only 0.24 s and 0.42 s. In addition, under the extreme conditions of 40% client drop rate or 50% data missing, the global model accuracy can still be stable, exceeding 98% on the MNIST dataset and 90% on the Fashion-MNIST dataset.