To　address　the　key　challenge　encountered　in　fake　news　detection,　i.e.,　multimodal　data　is　difficult　to　be　effectively　semantically　represented　due　to　its　intrinsic　heterogeneity,　this　paper　proposes　a　multimodal　knowledge　representation　method　for　fake　news　detection.　First,　visual　feature　extraction　is　performed　for　fake　news　image　data,　the　relevant　images　are　sliced　into　multiple　blocks,　and　then　visual　modal　features　are　obtained　by　linear　projection　layer　mapping.　This　simplifies　the　feature　extraction　process　and　reduces　the　computational　cost,　which　helps　to　improve　the　fake　news　recognition　performance.　Second,　to　meet　the　actual　fake　news　detection　needs,　a　long　text　representation　method　based　on　topic　words　is　investigated　for　the　text　data　in　fake　news.　Finally,　the　multimodal　representation　of　the　same　fake　news　data　is　optimized　by　establishing　a　connection　between　two　different　modalities,　visual　and　text,　and　inputting　it　into　a　BiLSTM-Attention　based　network　to　achieve　the　fusion　of　multimodal　features.　The　experiment　selects　the　same　fake　news　data　of　EANN　model　and　uses　four　classical　classification　methods　to　verify　the　effect　of　knowledge　representation　and　compare　it　with　the　fusion　model　ViLT　which　is　not　optimized　for　long　text.　The　experiment　proves　that　the　accuracy　rate　of　fake　news　detection　using　the　multimodal　representation　proposed　in　this　paper　is　improved　by　7.4%　compared　to　the　EANN　model,　and　by　9.3%　compared　to　the　ViLT　representation.