DeepPhish: Simulating Malicious AI
summary
Builds practical models that actually work: a char-level LSTM URL classifier trained on millions of URLs hits ~99% accuracy and runs on mobile; a certificate model trained on ~1M legit and ~5k phishing certs reaches ~86% accuracy; a DeepPhish pipeline clusters 1.1M confirmed phishing URLs into threat actors, trains a generator (Python, Keras, Jupyter) to synthesize path strings, and raises an attacker’s bypass rate from ~0.7% to 29% effectiveness (another actor to 36%), proving AI can amplify phishing success.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 7:07Recurrent Neural Network
- 9:28Web Certificate
- 12:43Long Name Hunting
- 18:42Research Project
- 23:51Data Analysis
- 28:01Algorithm Summary
- 30:20How to use Algorithm
- 31:56Effectiveness of Algorithm
- 33:42Results
- 35:24Attacks are real
- 40:04Phishing certificates
- 41:27Redirects
- 43:13Next paper
- 46:34Would I be able to render your service inefficient
- 48:21Does your algorithm update its web certificates
- 49:35Did you try to fish against stats
related talks
50mCatPhish Automation - The Emerging Use of Artificial Intelligence in Social Engineering
31mAI-Powered Solutions for Cybersecurity
21mMachine Learning and Cyber Security - Detecting malicious URLs in the haystack
11mHow to Navigate Emerging AI-Driven Security
49mDo Attackers Use Algorithms to Evade ML? An In-the-Wild Phishing Case Study
30mBuilding a Predictive Pipeline to Rapidly Detect Phishing Domains
05m[Flash talk] MUDPIPE - Malicious URL Detection for Phishing Identification and Prevention
01mCombatting scalable AI attacks