PhishJEPA – a Transferable Phishing Email-Related Representation Learning

Abstract Phishing is a primary initial vector for data breaches and results in high global costs. While deep learning models achieve near-perfect in-domain detection, cross-corpus distribution shifts severely degrade performance owing to overfitting on dataset-specific artefacts instead of learning semantic intent. To address this, we introduce PhishJEPA, a self-supervised Joint Embedding Predictive Architecture using Exponential Moving Average student-teacher updates and Variance-Invariance-Covariance Regularisation to model email subject-body semantic relationships in latent space. We systematically evaluate representation transferability across four unseen benchmark corpora in two stages over multiple random seeds. In Stage 1 (feature probing), JEPA representations demonstrate statistically robust improvement in cross-corpus separability over vanilla DistilBERT under a linear probe (+0.0124 pooled AUROC, p ~= 9x10 -11 ), whereas multilayer perceptron probing (p ~= 0.93)and unsupervised domain adaptation (-0.0039 AUROC, p~=1x10 -9 ) offer no gains. In Stage 2 (downstream classification), freezing the JEPA backbone and training a lightweight linear head achieves a cross-corpus performance (0.9394 AUROC / 0.9157 AUPRC) comparable to full end-to-end DistilBERT fine-tuning (0.9496 AUROC), while reducing the training time by 60% and peak-allocated GPU memory by 43%. Crucially, we identify an operational capacity tradeoff: higher-capacity MLP heads overfit in-domain (0.9980 AUROC) and suffer larger cross-corpus performance drops (7.24%-19.20%) than linear heads (5.19%-10.63%), proving far less threshold stability. Finally, post-processing per-domain isotonic calibration reduces the fixed false-positive rates by up to 16.7% across training sources without degrading ranking performance. These findings indicate that JEPA learns transferable semantic representations, but reliable deployment under distribution shifts depends critically on a simple classifier design and post-hoc calibration.