Using the data from the Large Area Telescope (LAT), the Fermi-LAT collaboration continuously updates their catalogs, which now contain a few thousands of detected gamma-ray sources. Among them, around one third are of not yet identified origin, and they could contain signals from established source types or, most intriguing, new source types such as dark matter subhalos producing gamma-rays from dark matter self-annihilation. We apply state-of-the-art machine learning methods for classification to the sources in Fermi-LAT catalogs with the aim of identifying possible candidates of exotic gamma-ray sources, namely dark matter subhalos. We first simulate the properties of dark matter subhalo gamma-ray sources by using established models from both N-body simulations and semi-analytical approaches for the subhalo distribution. We then carefully assess the detectability of this sample by using Fermi-LAT simulations. We discuss results of our machine learning analysis performed on the unidentified sources in the 4FGL-DR3, and present conservative limits on the dark matter annihilation cross-section from the exclusion of the unidentified sources classified as astrophysical-like by our networks.