Federated　Learning　(FL)　lately　has　shown　much　promise　in　improving　sharing　model　and　preserving　data　privacy.　However,　these　existing　methods　are　only　of　limited　utility　in　the　Internet　of　Things　(IoT)　scenarios,　as　they　either　heavily　depend　on　high-quality　labeled　data　or　only　perform　well　under　idealized　distribution　conditions,　which　typically　cannot　be　found　in　practical　applications.　In　this　work,　we　propose　FedGAN,　a　Generative　Adversarial　Network　(GAN)　based　federated　learning　method　for　semi-supervised　image　classification.　In　IoT　scenarios,　a　big　challenge　is　that　decentralized　data　among　multiple　clients　are　normally　non-independent　and　identically　distributed　(non-IID),　leading　to　performance　degradation.　To　address　this　issue,　we　further　propose　a　dynamic　aggregation　mechanism　that　can　adaptively　adjust　client　weights　in　aggregation.　Extensive　experiments　on　three　benchmarks　demonstrate　that　FedGAN　outperforms　related　federated　semi-supervised　learning　methods,　including　a　55.36%　accuracy　on　CIFAR-10　with　2k　labels　and　70.65%　accuracy　on　SVHN　with　1k　labels　-　just　100　labels　per　class.　Moreover,　we　carry　out　an　extensive　ablation　and　robust　study　to　tease　apart　the　experimental　factors　that　are　important　to　FedGAN＇s　improvement.