<p>Automatic Speaker Verification (ASV) systems are widely utilized in real-time settings, to ensure security in highly sensitive environments, these systems must be robust against known or unknown logical access attacks. Various convolutional neural network (CNN) architectures and their variants, along with advanced attention mechanisms, have been extensively studied in order to develop effective countermeasures for ASV systems. Despite these advancements, the challenge arises when detection systems encounter unfamiliar spoofing algorithms, making it more difficult to identify the most relevant regions for analysis. This often results in poor generalization capabilities. Notably, there are significant differences in the relationships between regions in genuine versus spoofed speech. By emphasizing overall dependency patterns instead of focusing solely on specific areas, detection systems can enhance their generalization performance. In this context, we propose a novel approach using Convolutional Kolmogorov-Arnold Networks (CKANs) to extract meaningful information for spoof detection tasks. The Proposed methodology is designed to capture dependencies effectively and integrate prior knowledge through careful structuring of the spectrograms of genuine and spoofed speech. By directing the network to prioritize potential relationships, our proposed method achieved a minimum Detection Cost Function (minDCF) result of 0.4161 on ASV Spoof 2021 LA dataset. It outperforms baseline systems (B1, B2 and B4) referred in ASVspoof 2021 Challenge. The outcomes show the potential capabilities of proposed novel CKAN network in the development of synthetic voice detection systems, highlighting the effectiveness of KAN-based architectures. Future research can further refine these models to develop a state-of-the-art synthetic voice detection system, ultimately strengthening the security of ASV systems.</p>