Exact Identity Aware Evaluation of Static Malware Images and API Sequences Using Multi Scale Features and Confidence Aware Selective Prediction

Duplicate artifacts can make malware classification results appear more general than they are.This study examines two representation specific cases after exact identity auditing.In the static case, removing 810 same label duplicates leaves 7,703 images across 30 classes.A 1,272 dimensional multi scale descriptor with a radial basis function support vector machine achieves 97.75 percent accuracy on 1,155 held out images.In the application programming interface case, 8,513 rows collapse to 499 exact identities, and six families provide 448 eligible unique sequences after conflict removal.A four view term frequency inverse document frequency linear support vector machine achieves 94.44 percent accuracy at full coverage.Confidence calibration increases selective accuracy to 98.6 percent at 77.78 percent coverage.Controlled reruns of earlier deep baselines on the cleaned partitions reach 97.40 percent static accuracy and 91.11 percent sequence accuracy.The results support identity controlled evaluation rather than multimodal fusion or state of the art superiority.