A　large-scale　model　is　typically　trained　on　an　extensive　dataset　to　update　its　parameters　and　enhance　its　classification　capabilities.　However,　directly　using　such　data　can　raise　significant　privacy　concerns,　especially　in　the　medical　field,　where　datasets　often　contain　sensitive　patient　information.　Federated　Learning　(FL)　offers　a　solution　by　enabling　multiple　parties　to　collaboratively　train　a　high-performance　model　without　sharing　their　raw　data.　Despite　this,　during　the　federated　training　process,　attackers　can　still　potentially　extract　private　information　from　local　models.　To　bolster　privacy　protections,　Differential　Privacy　(DP)　has　been　introduced　to　FL,　providing　stringent　safeguards.　However,　the　combination　of　DP　and　data　heterogeneity　can　often　lead　to　reduced　model　accuracy.　To　tackle　these　challenges,　we　introduce　a　sampling-memory　mechanism,　FedSam,　which　improves　the　accuracy　of　the　global　model　while　maintaining　the　required　noise　levels　for　differential　privacy.　This　mechanism　also　mitigates　the　adverse　effects　of　data　heterogeneity　in　heterogeneous　federated　environments,　thereby　improving　the　global　model＇s　overall　performance.　Experimental　evaluations　on　datasets　demonstrate　the　superiority　of　our　approach.　FedSam　achieves　a　classification　accuracy　of　95.03%,　significantly　outperforming　traditional　DP-FedAvg　(91.74%)　under　the　same　privacy　constraints,　highlighting　FedSam＇s　robustness　and　efficiency.　©　2025　Elsevier　B.V.