Network　Intrusion　Detection　(NID)　is　an　important　defense　strategy　in　modern　networks　to　detect　malicious　activities　in　large-scale　cyberspace.　The　current　NID　methods　suffer　from　a　high　false　positive　rate,　which　significantly　reduces　the　overall　effectiveness　of　network　intrusion　detection　systems　and　simultaneously　increases　the　maintenance　cost.　Furthermore,　the　class　imbalance　problem　associated　with　the　intrusion　detection　dataset　limits　the　detection　rate　for　the　minority　classes.　This　paper　proposes　a　novel　hierarchical　CNN-Attention　network,　CANET.　In　CANET,　CNN　and　the　Attention　mechanism　mingle　to　form　a　CA　Block　that　focuses　on　local　spatio-temporal　feature　extraction.　The　multi-layer　CA　Block　combination　can　fully　learn　the　multi-level　spatio-temporal　features　of　network　attack　data,　which　is　more　suitable　for　modern　large-scale　NID.　Besides,　for　the　class　imbalance　problem,　we　propose　to　use　Equalization　Loss　v2　(EQL　v2)　to　increase　the　minority　class　weight　and　balance　the　learning　attention　on　minority　classes.　Extensive　experiments　demonstrate　that　CANET　outperforms　the　state-of-the-art　methods　in　terms　of　accuracy,　detection　rate,　and　false　positive　rate.　And　it　efficiently　improves　the　detection　rate　of　minority　classes.　The　source　code　for　the　proposed　CANET　models　is　publicly　available　at　https://github.com/yuanshuai666/CANET.　©　2023　Elsevier　B.V.