Vulnerability and Mitigations in Machine Learning

1. Vulnerability

• CERT

  • Vulnerability Note VU#425163:
    Machine learning classifiers trained via gradient descent are vulnerable to arbitrary misclassification attack

• Experimental Security Research of Tesla Autopilot

  • Web
  • PDF

• Phantom of the ADAS: Phantom Attacks on Driver-Assistance Systems

  • Web
  • PDF

2. Mitigations

2.1 Adversarial training and testing of ML models

• Cleverhans
  • Blog
  • GitHub
• Foolbox
  • GitHub
• Adversarial Robustness Toolkit (ART)
  • GitHub

2.2 Standard defense in depth

• Machine Learning in Cybersecurity: A Guide
  • Web
  • PDF

3. Acknowledgements

• Towards the Science of Security and Privacy in Machine Learning
  • arXiv
• Wild Patterns: Ten Years After the Rise of Adversarial Machine Learning
  • arXiv
• On Adaptive Attacks to Adversarial Example Defenses
  • arXiv