KAN-Steganalysis: A Kolmogorov–Arnold Network Framework for Interpretable and Efficient Image Steganalysis

Hiding information in a carrier is called Steganography, whereas the process of detecting the presence of hidden information within any carrier such as digital images, audio, text is known as Steganalysis. In the recent decade the advent of Convolutional Neural Networks(CNN) has boosted the performance of Steganalysis techniques to higher level and eliminated the feature selection phase which requires considerable domain expertise. But in spite of better performance than traditional machine learning based Steganalysis, CNN based techniques have disadvantages like, huge number of parameters and moreover their performance cannot be interpreted. These are major problems in the case of limited data and subtle stego noise. In this paper we discuss and assess the applicability of Kolmogorov-Arnold Networks(KANs) for the task of Steganalysis instead of Convolutional Neural Networks(CNN). KAN networks use learnable spline functions instead of fixed point wise activation used in CNNs. KANs also offer higher interpretability The KolmogorovArnold Networks architecture consists of high-pass aware spline activations, adaptive channel gating mechanism and curvature regularization. As a result the number of parameters decreases compared to CNN. BOSSBase 1.01 dataset is used in our work and the detection accuracy of KAN based Steganalysis is found to be on par with the contemporary CNN based Steganalysis techniques such as SRNet and Yedroudj-Net with requirement of 40% less parameters. This spline based representation leads to interpretable activation patterns with which we can have the insight into the decision process of the Steganalysis model. These results establish the potential of functional neural networks as an efficient and transparent alternative for modern image Steganalysis.