Directed　transfer　function　(DTF)　is　good　at　characterizing　the　pairwise　interactions　from　whole　brain　network　and　has　been　applied　in　discrimination　of　motor　imagery　(MI)　tasks.　Considering　the　fact　that　MI　electroencephalogram　signals　are　more　non-stationary　in　frequency　domain　than　in　time　domain,　and　the　activated　intensities　of　α　band　(8–13 Hz)　and　β　band　[13–30 Hz,　with　β1(13–21 Hz)　and　β2(21–30 Hz)　included]　have　considerable　differences　for　different　subjects,　a　dynamic　DTF　(DDTF)　with　variable　model　order　and　frequency　band　is　proposed　to　construct　the　brain　functional　networks　(BFNs),　whose　information　flows　and　outflows　are　further　calculated　as　network　features　and　evaluated　by　support　vector　machine.　Extensive　experiments　are　conducted　based　on　a　public　BCI　competition　dataset　and　a　real-world　dataset,　the　highest　recognition　rate　achieve　100%　and　86%,　respectively.　The　experimental　results　suggest　that　DDTF　can　reflect　the　dynamic　evolution　of　BFN,　the　best　subject-based　DDTF　appears　in　one　of　four　frequency　sub-bands　(α,　β,　β1,　β2)　for　discrimination　of　MI　tasks　and　is　much　more　related　to　the　current　and　previous　states.　Besides,　DDTF　is　superior　compared　to　granger　causality-based　and　traditional　feature　extraction　methods,　the　t-test　and　Kappa　values　show　its　statistical　significance　and　high　consistency　as　well.　©　2022,　The　Author(s).