Due　to　technical　barriers　and　economic　costs,　malicious　circuits,　known　as　hardware　Trojans,　are　easily　implanted　in　the　complicated　integrated　circuit　design　and　manufacturing　process,　which　can　lead　to　many　disastrous　consequences,　such　as　denial　of　service,　information　leakage,　performance　degradation,　etc.　Research　on　how　to　detecting　hardware　Trojans　has　grown　into　a　significantly　open　issue　over　the　past　decade.　While,　for　very　large　scale　integrated　circuits,　numerous　new　challenges　deserve　our　full　attention,　including　golden-free　chip　reference,　automatic　feature　engineering,　hardware　Trojan　localization,　and　scalable　framework.　In　response　to　the　above　challenges,　a　fine-grained　gate-level　hardware　Trojan　detection　approach　is　proposed　in　this　paper,　named　GateDet,　from　improving　earlier　circuit　graph　modeling　to　developing　a　detection　framework　based　on　Bidirectional　Graph　Convolution　Networks　with　a　timely　information　fusion　strategy.　GateDet　achieves　automatic　feature　circuit　extraction　and　further　overcomes　the　original　neighborhood　limitation　of　Bidirectional　Graph　Convolution　Network.　Moreover,　for　large-scale　training,　it　comprehensively　considers　the　problems　of　sample　imbalance　and　boundary　network,　and　develops　a　circuit　directed　graph　sampling　method　based　on　GraphSAINT,　which　improves　the　training　performance　of　the　directed　graph　framework.　From　experiments,　GateDet　shows　high　scalability　on　24　benchmarks　of　TrustHub.　It　could　be　used　to　learn　about　adaptive　structural　feature　extraction　for　different　Trojans　simultaneously.　Compared　to　the　existing　gate-level　detections,　the　fine-grained　results　of　GateDet　are　more　accurate　and　can　be　used　to　track　suspicious　structures,　reducing　manual　review.　©　2023　The　Author(s)