The　application　of　federated　learning　in　training　brain　tumor　classification　models　not　only　safeguards　the　privacy　of　medical　data　but　also　facilitates　collaboration　among　multiple　healthcare　providers.　Nonetheless,　many　existing　federated　learning　algorithms　for　brain　tumor　diagnosis　heavily　lean　on　traditional　deep　learning　concepts,　often　disregarding　the　challenges　presented　by　the　distribution　of　non-IID　data　in　the　federated　setting.　This　research　introduces　a　federated　learning　(FL)　algorithm　that　employs　server　model　perturbation　to　address　the　impacts　of　non-IID　data　in　collaborative　learning.　Furthermore,　a　partial　batch　normalization　technique　is　implemented　to　improve　the　model’s　performance　in　non-IID　environments.　Additionally,　a　method　is　proposed　in　this　study　that　effectively　combines　these　two　approaches　to　enhance　the　algorithm’s　accuracy　in　brain　tumor　classification.　To　validate　this　methodology,　experiments　were　conducted　using　publicly　available　non-IID　brain　tumor　processing　datasets,　comparing　the　approach　with　cutting-edge　collaborative　learning　algorithms　like　FedAvg,　FedProx,　and　FedBN.　The　results　clearly　indicate　the　exceptional　effectiveness　of　the　proposed　approach,　showing　an　average　improvement　of　about　1%　in　the　statistical　heterogeneity　setting　and　approximately　3%　in　the　feature　heterogeneity　setting.　©　2024　Copyright　held　by　the　owner/author(s).