Breast cancer histopathological image classification plays a vital role in the early diagnosis and treatment strategy. The data must be stored centrally for the traditional machine learning approach, which may raise privacy concerns and restrict cross-institutional collaboration. Federated learning may solve these problems by facilitating model training at the edge while preserving patient data privacy. The approach outlined in this work follows strict data-sharing protocols and thus is suitable for health uses. We propose a federated learning framework suitable for histopathological image classification in this work. Our approach utilizes secure data processing and model aggregation protocols to reach data confidentiality as well as effective updating of the models. We validate experimentally that our model has better classification accuracy than the traditional centralized models. In addition, it enables interaction between multiple medical centers without sacrificing confidential information. The experiment proves the efficiency of federated learning in better breast cancer diagnosis with AI for privacy protection. More sophisticated aggregation algorithms and domain adaptation can be further studied in subsequent research to improve the accuracy of classification across different clinical environments.