The　bulk　of　existing　Federated　Learning　(FL)　algorithms　pay　attention　to　supervised　setting　and　assume　that　clients　have　fully　labeled　data.　However,　it　may　be　impractical　for　all　clients　to　obtain　plenty　of　labels　due　to　high　annotation　costs.　Hence,　the　Federated　Semi-Supervised　Learning　(FSSL)　as　a　promising　paradigm　has　better　prospect　in　many　realistic　applications　(e.g.　medical　scenario).　Under　Labels-at-Server　scenario,　most　pseudo　labeling　based　FSSL　approaches　use　only　the　global　model　to　generate　pseudo-labels　for　unlabeled　data,　while　the　local　models　are　ignored.　When　the　local　data　distribution　is　much　more　different　from　the　central　server　(e.g.,　Non-IID　setting),　the　generated　pseudo-labels　may　contain　much　noise,　thus,　resulting　in　more　serious　confirmation　bias.　To　tackle　the　crucial　issue,　a　novel　Federated　Semi-Supervised　Learning　via　Joint　Local　and　Global　Pseudo　Labeling　(FedLGMatch)　framework　is　proposed　in　this　paper.　The　prominent　advantage　of　the　proposed　FedLGMatch　is　that　it　allows　local　models　trained　in　the　last　communication　round　to　assist　global　model　in　generating　pseudo-labels,　which　successfully　emphasizes　more　clean　pseudo-label　learning　at　each　client.　Experimental　results　also　show　that　FedLGMatch　achieves　significant　performance　improvements　than　other　state-of-the-art　models　on　the　standard　benchmark　datasets.