Distributed　learning　applied　in　industrial　cyber-physical　systems　(ICPS)　is　vulnerable　to　cyber　attacks,　especially　rare　ones.　Common　data-driven　cyber　attack　detection　approaches　face　the　challenges　of　imbalanced　data,　resulting　in　insufficient　extraction　of　anomalous　features.　To　enhance　the　sensitivity　of　rare　cyber　attack　detection　in　complex　ICPS,　we　propose　a　federated　diffusion-squeeze　graph　model　(FedDSG).　In　each　edge　device,　we　construct　a　local　diffusion-based　generative　module　to　balance　rare　anomalous　data　and　construct　feature　graphs,　which　maintains　information　fidelity　and　type　balance　of　data.　To　alleviate　the　extra　computational　load,　we　establish　a　graph-structured　detection　module　based　on　information　bottleneck　(IB)　to　filter　out　redundant　topological　features　and　identify　the　optimal　graph　for　modeling.　In　the　central　server,　we　design　an　aggregation　strategy　in　the　central　server　to　consolidate　a　global　FedDSG　and　the　global　generative　module　generates　synthetic　cyber　attack　data　to　retrain　the　global　detection　module.　In　addition,　we　verify　FedDSG　using　public　industrial　datasets　on　the　self-constructed　simulation　platform.　The　results　show　that　FedDSG　improves　the　efficiency　of　rare　cyber　attack　detection.　©　2025　IEEE.　All　rights　reserved,