Federated　learning　addresses　the　＇data　islands＇　issue　resulting　from　privacy　concerns,　enabling　multiple　users　to　train　a　model　collaboratively　without　sharing　their　private　datasets.　However,　in　the　model　training　process　of　federated　learning,　effectively　coordinating　the　communication　between　participants　and　server　is　essential　to　improve　the　training　efficiency.　Traditional　synchronous　and　asynchronous　federated　learning　algorithms　cannot　meet　the　requirements　of　convergence　efficiency　and　precision　stability　at　the　same　time.　In　order　to　balance　algorithm　efficiency　and　model　stability,　this　paper　proposes　a　model　parameter　update　strategy　for　an　enhanced　asynchronous　federated　learning　algorithm　(EAFL).　This　strategy　requires　the　server　to　store　the　latest　global　model　parameters,　the　last　round　of　global　model　parameters,　and　the　update　part　of　both,　and　judge　the　issued　model　type　according　to　the　sensitivity　difference　of　the　client.　The　client　executes　different　local　update　processes　according　to　different　model　types　to　improve　the　training　efficiency　and　maintain　the　stability　of　the　training　accuracy.　The　experimental　results　show　that　the　EAFL　algorithm　achieves　better　training　results　on　MNIST　and　CIFAR-10　data　sets　than　the　traditional　federated　learning　parameter　update　algorithm.　This　improvement　in　training　results　can　provide　new　thoughts　to　solve　computer　vision　tasks　such　as　object　detection　with　scattered　devices.　©　2023　IEEE.