With　the　widespread　deployment　of　Power　Internet　of　Things　(PIoT),　wireless　communication　security　faces　increasingly　critical　challenges　from　eavesdropping　attacks.　Traditional　methods　for　localizing　eavesdropping　nodes　struggles　to　fully　leverage　network　structural　information,　often　result　in　high　model　complexity　and　limited　practicality.　To　address　these　issues,　this　paper　proposes　an　innovative　Graph　Convolutional　Attention　(GCAT)　localization　algorithm.　This　method　utilizes　graph　convolution　to　capture　the　global　topological　features　of　the　network　while　introducing　an　attention　mechanism　to　adaptively　aggregate　key　regions,　thus　enabling　a　fine-grained　spatiotemporal　representation　of　eavesdropping　behavior.　Extensive　simulation　experiments　show　that　GCAT　significantly　outperforms　traditional　methods　in　terms　of　accuracy,　recall,　and　other　metrics,　providing　new　insights　for　securing　PIoT　communication.　Notably,　GCAT　is　able　to　maintain　stable　performance　at　extremely　low　labeling　rates,　which　are　less　than　5%,　demonstrating　excellent　few-shot　learning　ability.　Simultaneously,　its　inference　latency　can　be　controlled　within　30　ms,　and　the　number　of　parameters　is　reduced　by　more　than　30%　compared　to　mainstream　GNN　models,　making　it　easy　to　deploy　in　resource-constrained　environments.　The　research　results　of　this　paper　can　provide　crucial　technical　support　for　the　construction　of　an　active　defense　system　for　PIoT,　and　have　significant　implications　for　ensuring　the　cybersecurity　of　energy　systems,　and　promoting　the　safe　and　controllable　development　of　ubiquitous　PIoT　applications　and　smart　grids.　　©　2025　IEEE.