JA4-JEPA applies JEPA-style predictive learning to JA4 network fingerprints

The authors explore whether JEPA-style predictive learning objectives, which match latent predictions to target encoder outputs, can be effectively applied to compact network fingerprints. They introduce JA4-JEPA, a Transformer-based model trained on approximately 397K samples of JA4, JA4H, JA4S, and JA4X subfields from JA4DB and CIC-IDS-2017 datasets.

• The model was evaluated using a frozen kNN probe for protocol-family classification across TLS, DNS, and SSH.
• On 39,416 held-out 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 when there is incomplete view overlap across training sources.