With　the　development　of　the　Internet,　security　issues　in　the　network　have　attracted　more　and　more　attention.　Variants　of　malicious　code　are　constantly　increasing,　and　their　attacks　will　have　a　serious　impact　on　the　network　environment,　so　effective　detection　of　malicious　code　has　important　research　significance.　However,　the　current　malicious　code　detection　methods　still　have　some　problems,　such　as　code　detection,　cumbersome　feature　extraction,　and　misclassification　between　similar　families.　To　this　end,　the　paper　proposes　a　fine-grained　detection　method　for　malicious　code.　First　visualized　the　binary　files　of　malicious　code　and　converted　them　into　grayscale　images.　Then,　use　the　improved　convolutional　neural　network　to　extract　the　multi-resolution　features　of　grayscale　images,　and　use　the　interactive　fusion　method　to　fuse　these　features.　Finally,　input　the　fused　features　into　the　fully　connected　layer　to　complete　the　fine-grained　classification　of　malicious　code.　Experiments　prove　that　our　method　is　indeed　effective　for　fine-grained　classification　of　malicious　code.　©　2021　IEEE.