In this research, the problem of blind detection of Polarization Shift Keying (POLSK) modulation is investigated in the presence of Additive White Gaussian Noise (AWGN) and alongside other common satellite communication modulations (PSK and QAM). Given the critical importance of security and anti-jamming capabilities in satellite links, POLSK emerges as a novel alternative capable of embedding information within the polarization domain. To this end, Stokes vectors extracted from the received signals were utilized as input features for two deep learning architectures: a Multilayer Perceptron (MLP) and a Quaternion Neural Network (QNN). The detection performance of these models was evaluated based on classification accuracy and confusion matrices over a SNR range of 0 to 20 dB. For a comprehensive evaluation, a large-scale dataset comprising approximately six million Stokes vector samples per class was generated. Simulation results on the combined dataset—including POLSK, PSK, QAM, and noise—demonstrate that the QNN significantly outperforms the MLP by leveraging quaternion algebra and effectively modeling the multidimensional correlations among the Stokes vector components. The QNN architecture achieved an average overall accuracy of 95.8% across the entire SNR range, substantially minimizing classification errors, particularly in isolating POLSK from other modulations. Furthermore, analysis of the confusion matrices reveals that the QNN, unlike the PMLP, is highly capable of discriminating the POLSK class from interference and noise, establishing it as a highly robust and suitable candidate for blind detection in practical satellite communication applications.