---
license: mit
tags:
- text-classification
- phishing-detection
- distilbert
- adversarial-robustness
---

# DistilBERT Phishing Classifier — SMS Spam

Fine-tuned `distilbert-base-uncased` for binary phishing/spam classification,
trained on the **SMS Spam** 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)