DeepPhish: Simulating Malicious AI
summary
A char-level RNN/LSTM model runs on URLs and reaches about 98–99% accuracy on ~2 million samples, implemented in Python/Keras and runnable on mobile; a second deep model flags phishing TLS certificates with ~86% accuracy from ~1M legit and ~5k phishing certs; a generator (LSTM predicting next characters) trains on the attacker’s successful URL patterns and produces new paths for known compromised domains, boosting an attacker’s effectiveness from 0.7% to ~29% (another actor from 5.9% to 36%); code and notebooks are on GitHub under a Jupyter workflow char-level RNN/LSTM.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 7:08Recurrent Neural Network
- 9:28Web Certificates
- 12:43Long Name Hunting
- 23:13Data Analysis
- 30:18URL Prediction
- 35:23Attacks are real
- 40:03Phishing certificates
- 41:26Directs
- 43:12Retraining
- 44:22False positives
- 46:33Would I be able to render your service inefficient
- 48:20Does your algorithm update its web certificates
- 49:34Are you aware of tools similar to Saltfish
related talks
31mAI-Powered Solutions for Cybersecurity
21mMachine Learning and Cyber Security - Detecting malicious URLs in the haystack
50mCatPhish Automation - The Emerging Use of Artificial Intelligence in Social Engineering
30mBuilding a Predictive Pipeline to Rapidly Detect Phishing Domains
49mDo Attackers Use Algorithms to Evade ML? An In-the-Wild Phishing Case Study
58mCatching Transparent Phish: Understanding and Detecting MITM Phishing Kits
11mHow to Navigate Emerging AI-Driven Security
38mPhishing LLMs: Reeling in the Machine