Weight-Stable Selection of Vision Backbones for Malware Classification Across Binary Obfuscation Conditions

초록

Image-based malware classifiers are commonly compared by choosing the architecture with the largest accuracy or F1 value in a single file condition. That practice can conceal a deployment-relevant conflict: an architecture may lead on unobfuscated executables yet lose precision after obfuscation, or may appear preferable only under a narrowly chosen importance pattern. This study asks which of four convolutional backbones—ResNet18, ResNet34, EfficientNetB3, and EfficientNetV2S—provides the most defensible choice when accuracy, precision, recall, and F1 are considered jointly for unobfuscated binaries and for binaries processed with XOR or Shikata Ga Nai. A condition-stratified decision matrix containing 32 performance values is examined through three complementary operations: coordinatewise dominance, floor- and-loss robustness analysis, and integration over uncertain criterion weights. The weighting operation samples one million vectors uniformly from the eight-dimensional probability simplex and records the first-ranked architecture for every vector. ResNet18 has the highest overall arithmetic mean (98.420%), the highest harmonic mean (98.402%), and a performance floor of 95.650%. It defeats ResNet34 on all eight coordinates and exceeds EfficientNetB3 on seven of eight. ResNet18 is selected for 99.9769% of the sampled weight vectors; EfficientNetB3 receives the remaining 0.0231%, owing to its 95.72% precision on obfuscated binaries. EfficientNetB3 nevertheless has the smallest average decline under obfuscation (1.4825 percentage points) and the highest floor (95.72%). The findings answer the selection question conditionally: ResNet18 is the weight-stable default, whereas EfficientNetB3 is justified only when obfuscated-file precision or minimum-coordinate protection is assigned exceptional priority. The analysis replaces single-number ranking with an auditable decision statement that separates overall superiority, resistance to condition shift, and preference sensitivity. © 2026 Societe de Physique et d'Histoire Naturelle de Geneve. All rights reserved.

키워드