Video　processing　has　consistently　been　a　focal　point　of　interest　in　both　the　media　industry　and　academia,　particularly　against　the　backdrop　of　continuous　advancements　in　digital　signal　processing　technology.　These　videos　are　sometimes　misused　to　disseminate　false　information　or　damage　reputations.　This　paper　aims　to　develop　a　video　authenticity　recognition　system　using　advanced　deep　learning　models.　In　terms　of　technical　implementation,　the　ResNet50　model　is　employed　for　image　feature　extraction　to　detect　subtle　inconsistencies　and　artifacts,　such　as　unnatural　facial　movements　and　irregular　lighting　variations　in　videos.　Additionally,　LSTM　is　used　to　capture　the　temporal　sequence　features　of　the　video.　Experimental　results　indicate　that　the　model　achieves　an　accuracy　rate　of　81.25%　in　detecting　deepfake　videos.　Additionally,　the　model　achieved　a　recall　rate　of　1,　which　holds　significant　practical　value.　This　achievement　provides　strong　support　for　the　accurate　detection　and　labeling　of　videos,　playing　a　crucial　role　in　preserving　information　authenticity　and　ensuring　media　security.　©　2025　IEEE.