STEGEFFICIENTNETB0: A COMPOUND-SCALED EFFICIENTNETB0 FRAMEWORK FOR SPATIAL-DOMAIN IMAGE STEGANALYSIS AND ANOMALY LOCALIZATION

Abstract

This study presents StegEfficientNetB0, a spatial-domain steganalyzer that leverages the compound scaling mechanism of EfficientNetB0 to enhance detection accuracy and computational efficiency. The model employs transfer learning from an ImageNet-pretrained backbone and is fine-tuned to distinguish between cover and stego images generated using five standard content-adaptive steganographic algorithms HUGO, WOW, HILL, MiPOD, and SUNIWARD at payloads of 0.2 and 0.4 bpp. By combining compound scaling and transfer learning, the proposed framework effectively captures subtle steganographic distortions across benchmark datasets BOSSBase1.01 and BOWS2. Experimental results demonstrate that StegEfficientNetB0 achieves 92.31% classification accuracy with approximately 4 million parameters, underscoring its efficiency and strong generalization capability. StegEfficientNetB0 outperforms existing CNN-based steganalyzers including YeNet, Yedroudj-Net, ZhuNet, SRNet, and GBRAS-Net, demonstrating superior accuracy with fewer parameters. The model’s balanced precision and recall across both classes validate its robustness, establishing StegEfficientNetB0 as a compact yet powerful framework for modern spatial steganalysis.