Epilepsy is a chronic neurological disorder characterized by recurrent seizures due to abnormal brain activity. Accurate detection of seizures from electroencephalogram (EEG) signals is critical, but it is often challenged by signal noise and class imbalance in real-world data. In this study, we systematically evaluate Kolmogorov–Arnold Networks (KANs)—a recent neural architecture based on the Kolmogorov–Arnold representation theorem—as an alternative to Multi-Layer Perceptrons (MLPs) for EEG-based seizure classification, with a focus on model robustness under noisy conditions. This is the first comprehensive evaluation of KAN's robustness under multiplicative noise in the context of EEG seizure detection. Experiments were conducted using two widely used EEG datasets: the Bonn dataset and the CHB-MIT Scalp EEG dataset. Across multiple network configurations and varying levels of multiplicative noise, we assess performance using F1 Score, AUROC, AUPRC, Sensitivity, and Specificity. Our findings show that KAN achieves more stable performance than MLPs under noisy conditions, particularly in smaller architectures. These results suggest that KAN may offer a robust and generalizable approach for seizure detection in noise-prone clinical settings.