JEPA-Style Learning Creates Useful Network Fingerprint Embeddings

The I-JEPA and V-JEPA models have shown success in learning by predicting latent representations rather than regenerating full inputs, particularly for image and video data. This research explores whether this predictive learning objective can be effectively applied to compact network fingerprints. A new model, JA4-JEPA, was developed using a Transformer architecture and trained on various JA4 subfields from different data sources. The training dataset combined a significant number of samples, even though no single sample contained all four view families of fingerprints. The model's learned representations were evaluated using a kNN probe for classifying protocol families (TLS, DNS, SSH). On a held-out dataset, JA4-JEPA achieved a high cosine similarity and kNN accuracy, demonstrating that JEPA-style predictive learning can indeed generate valuable embeddings from JA4-derived network fingerprints, even when the input data has incomplete views.