Phisher vs Filter
My thesis: does AI help phishers or spam filters more? I'm measuring it.
Thesis, in progressResearch
Spam filters and the people trying to get phishing emails past them have been in an arms race for twenty years, and both sides now have AI. Plenty of people have an opinion on which side it helps more; my thesis puts a number on it.
I build an attacker and a defender at four levels of sophistication and ask how much each step up in technology buys each side. Every attacker faces every defender, then the strongest pair fight over ten rounds while both keep learning. Snapshots from every round are played against each other afterwards, which shows whether either side really improved or is just running on a treadmill.
The twist is that the defender plays with a handicap. A filter that gets too aggressive blocks real mail, which for a business is often worse than letting a scam through, so each defender is also scored under a fixed limit on false positives.
The result is a single offence-defence balance: above 1 means AI favours attackers, below 1 means it favours defenders. It runs entirely offline, no email can ever be sent, and the generated attack emails won't be released.
Final-year thesis, approved by my supervisor. Results to come.
What's in it
Level 0: no AI
Attacker: synonym swaps, look-alike characters and padding. Defender: hand-written SpamAssassin rules.
Level 1: classic ML
Attacker: a simple statistical model picks which words to change. Defender: a traditional spam classifier.
Level 2: neural
Attacker: a BERT language model suggests replacement words. Defender: a BERT classifier, trained once.
Level 3: generative and adaptive
Attacker: a small local LLM rewrites the whole email. Defender: a BERT classifier that keeps learning from its own misses.
Built with
- Python
- PyTorch
- Hugging Face
- scikit-learn
- SpamAssassin
- LoRA