In this poster, we start exploring whether ML-based methods can extract meaningful structural properties from a set of Internet packets originating from different sources. Unlike prior work, we aim for a technique that requires neither pretraining, labeling nor supervision, and can be executed locally, enabling sovereign and reproducible research. As a first approach, we combine contrastive learning and correlation clustering to identify groups of Internet scanners. Our analysis is based on data from the UCSD network telescope.