Asian Journal of Mathematics and Computer Research

Image steganalysis aims to determine whether digital images contain covertly embedded information. Adaptive steganographic methods can distribute modifications across image regions with different texture characteristics, making weak embedding-related signals difficult to detect using convolutional kernels with fixed receptive fields. This study evaluates an image steganalysis model that incorporates Selective Kernel (SK) attention into the SiaStegNet architecture. The proposed model retains a Siamese feature-extraction structure while replacing selected convolutional blocks with SK attention blocks that adaptively weight features obtained from different receptive fields. Evaluation was conducted using the BOSSBase 1.01 dataset, with images cropped to 256 × 256 pixels and stego images generated using WOW, HILL and S-UNIWARD at embedding payloads of 0.1, 0.2, 0.3 and 0.4 bits per pixel. Detection accuracy, area under the receiver operating characteristic curve, model parameter count and training duration were compared with those of the original SiaStegNet. For WOW at 0.4 bits per pixel, accuracy increased from 90.77% to 92.47%, and AUC increased from 97.82% to 98.50%. The proposed model used 0.46 million parameters compared with 0.71 million in the baseline, with the reported time per epoch decreasing from 115 to 100 seconds. Although improvements were observed in most tested configurations, accuracy was not higher in every case. The results indicate that SK attention can improve detection performance and computational efficiency within the evaluated setting.