DistilBERT Phishing Classifier — Spam Detection
Fine-tuned distilbert-base-uncased for binary phishing/spam classification,
trained on the Spam Detection dataset as part of an empirical study on
adversarial robustness (homoglyph substitution vs. synonym replacement attacks).
- Training: 3 epochs, batch size 32, AdamW, lr=2e-5
- Baseline accuracy: see paper Table/Fig. 1
- Part of: "Adversarial Robustness of NLP-Based Phishing Email Classifiers:
A Multi-Dataset Empirical Study" (CJSJ 2026, Muktadir Arif)
