Abstract
I-JEPA and V-JEPA learn by matching latent predictions to target encoder outputs rather than regenerating the original input, achieving success particularly with images and videos. We explore whether this objective can be applied to compact network fingerprints.
We built JA4-JEPA, a Transformer-based model trained on the JA4, JA4H, JA4S, and JA4X subfields drawn from JA4DB and CIC-IDS-2017. The training data combines roughly 397K samples from both sources, with no single sample containing all four view families.
We evaluated the learned representations with a frozen kNN probe on protocol-family classification across TLS, DNS, and SSH. On 39,416 heldout samples, the model achieved a cosine similarity of 0.9899 and a kNN accuracy of 0.9220. These results indicate that JEPA-style predictive learning can produce useful embeddings from JA4-derived fingerprints, even with incomplete view overlap across sources.
Blogger's Review: This study highlights the potential of the JEPA model in the domain of network fingerprinting, particularly under conditions of incomplete data samples. Its efficient learning capabilities provide new insights for network security and traffic analysis, potentially leading to innovative applications of self-supervised learning techniques in the cybersecurity field.