Machine Learning Testing Ecosystem of Python

  • Yunus Emrah Bulut

  • https://2022.pycon.de/program/9UB3Z3/

  • How to audit ML models?

  • Researchers have identified vulnerabilites in ML model.

Vulnerabilites of ML

  1. adversial attack
  2. Leaking private info
  3. Results are unexplainable
  4. Can be unfair in decision making

1. Evasion attack.

  • E.g. Tesla autopilot was fooled driving from 35kmph to 85 kmph
  • https://electrek.co/2020/02/19/tesla-autopilot-tricked-accelerate-speed-limit-sign/
  • A tape was used in Speed sign.

2. membership inference attack

3. Physical attack

4. Biased decision

  • Make biased decision from data they are trained from

5. Change in Data distribtuion

  • Relation between features and target changes
  • e.g. COVID period

Regulators

  • EU regulations has mandatory
  • validation for High Risk AI system (HRAI)

Tools:

  1. ART - adversirial robustness tool
  2. cleverbans
  3. Fairlearn: Python package

Validiator

  • New tool developed by the author and university