In　recent　years,　extensive　research　has　been　conducted　on　the　attack　and　defense　of　adversarial　examples　within　the　domain　of　deep　learning.　In　particular,　the　transferability　of　adversarial　examples　in　black-box　attack　scenarios　has　garnered　significant　attention.　The　generation　of　highly　transferable　adversarial　examples　presents　a　notable　challenge.　Traditional　approaches　involve　modifying　the　intermediate　layers　of　surrogate　model　networks　or　utilizing　data　augmentation　on　input　images　for　the　generation　of　transferable　adversarial　examples.　However,　generating　adversarial　examples　with　enhanced　transferability　poses　a　significant　challenge.　Thus,　in　this　paper,　we　proposes　a　novel　methodology　that　integrates　model　ensemble　and　a　distribution　perspective.　It　entails　initial　fine-Tuning　of　a　surrogate　model,　subsequent　utilization　of　the　fine-Tuned　surrogate　model　to　create　a　network　pool,　and　then　leveraging　this　network　pool　for　the　generation　of　adversarial　examples.　Subsequently,　these　generated　adversarial　examples　are　employed　to　attack　other　network　structures.　Empirical　findings　confirm　the　effectiveness　of　our　approach　in　achieving　a　notably　high　success　rate　in　adversarial　attacks.　©　2024　ACM.