Artificial intelligence (AI) applications in biomedical settings face challenges such as data privacy and regulatory compliance. Federated Deep Learning (FDL) effectively addresses these issues. We developed ProCanFDL, where local models were trained on simulated sites using proteomic data drawn from a pan-cancer cohort (n = 1,260) and 29 other cohorts (n = 6,265), representing 4,956 patients and 19,930 mass spectrometry (MS) runs, all held behind private firewalls. Local parameter updates were aggregated to build the global model, achieving a 43% performance gain over local models on the hold-out test set (n = 625) in 14 cancer subtyping tasks. Additionally, ProCanFDL preserved data privacy while matching centralized model performance. External validation assessed generalization by retraining the global model with data from two external cohorts (n = 55) and eight (n = 832) using a different MS technology. ProCanFDL presents a solution for internationally collaborative machine learning initiatives using proteomic data while maintaining data privacy.