Using Machines to exploit Machines
summary
Claims an ML pipeline that triages fuzzing crashes faster and more accurately than the heuristic tool exploitable, using features extracted from crash state (register values binned, instruction pointer, backtrace and input), a Docker harness and the DARPA Cyber Grand Challenge dataset of 632 labeled crashes as ground truth. Experiments use one‑class SVM, cosine similarity and XGBoost, finding up to 22 exploitable crashes missed by exploitable and nominal 95–99% accuracy but note heavy dataset bias and the need for nonexploitable controls before production use.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:32The problem: Triage
- 5:43Project origin
- 8:29What is machine learning?
- 13:45ML benefits and caveats
- 17:05Human vs machine intuition
- 21:51Data sources and collection
- 25:25Applying algorithms
- 31:43Reflecting on model accuracy
- 35:10Conclusion and call to action
- 40:28Q&A session
related talks
48mUsing Machine Learning to find Linux bugs
44mTL;DR: Applying AI to Security
24mTesting Techniques for Critical Deep Learning Applications
39mDetecting Website Intrusion and Account Compromise with Machine Learning
46mSW Testing: Can ML save us?
32mAI/Machine Learning: What is Actually Working in Cybersecurity?
10mML-based Log Analysis for Faster Debugging
43mThe Aftermath of a Fuzz Run: What to do about those Crashes?