Deep Learning-based computational pathology algorithms have demonstrated\nprofound ability to excel in a wide array of tasks that range from\ncharacterization of well known morphological phenotypes to predicting\nnon-human-identifiable features from histology such as molecular alterations.\nHowever, the development of robust, adaptable, and accurate deep learning-based\nmodels often rely on the collection and time-costly curation large high-quality\nannotated training data that should ideally come from diverse sources and\npatient populations to cater for the heterogeneity that exists in such\ndatasets. Multi-centric and collaborative integration of medical data across\nmultiple institutions can naturally help overcome this challenge and boost the\nmodel performance but is limited by privacy concerns amongst other difficulties\nthat may arise in the complex data sharing process as models scale towards\nusing hundreds of thousands of gigapixel whole slide images. In this paper, we\nintroduce privacy-preserving federated learning for gigapixel whole slide\nimages in computational pathology using weakly-supervised attention multiple\ninstance learning and differential privacy. We evaluated our approach on two\ndifferent diagnostic problems using thousands of histology whole slide images\nwith only slide-level labels. Additionally, we present a weakly-supervised\nlearning framework for survival prediction and patient stratification from\nwhole slide images and demonstrate its effectiveness in a federated setting.\nOur results show that using federated learning, we can effectively develop\naccurate weakly supervised deep learning models from distributed data silos\nwithout direct data sharing and its associated complexities, while also\npreserving differential privacy using randomized noise generation.\n