qwqoro/ml-talk

📄 [Talk] OFFZONE 2022 / ODS Data Halloween 2022: Black-box attacks on ML models + with use of open-source tools

[ Have a look at the presentation slides: slides-OFFZONE.pdf / slides-ODS.pdf ]
[ Related demonstration (Jupyter notebook): demo.ipynb ]

Overview | Attacks | Tools | More on the topic

An overview of black-box attacks on AI and tools that might be useful during security testing of machine learning models.

📦 Overview

demo.ipynb:
A demonstration of use of multifunctional tools during security testing of machine learning models digits_blackbox & digits_keras trained on the MNIST dataset and provided in Counterfit as example targets.

Slides:
 – Machine Learning in products
 – Threats to Machine Learning models
 – Example model overview
 – Evasion attacks
 – Model inversion attacks
 – Model extraction attacks
 – Defences
 – Adversarial Robustness Toolbox
 – Counterfit

⚔️ Attacks

• Model inversion attack: MIFace — code / docs / 🔗DOI:10.1145/2810103.2813677
• Model extraction attack: Copycat CNN — code / docs / 🔗arXiv:1806.05476
• Evasion attack: Fast Gradient Method (FGM) — code / docs / 🔗arXiv:1412.6572
• + Evasion attack: HopSkipJump — code / docs / 🔗arXiv:1904.02144
  

🔧 Tools

 – [ Trusted AI, IBM ] Adversarial Robustness Toolbox (ART): Trusted-AI/adversarial-robustness-toolbox
 – [ Microsoft Azure ] Counterfit: Azure/counterfit

📑 More on the topic

• adversarial examples evasion attacks
  How MIT researchers made Google's AI think tabby cat is guacamole: overview / 🔗arXiv:1707.07397 + arXiv:1804.08598

• model inversion attacks
  Apple's take on model inversion: overview / 🔗arXiv:2111.03702

• model inversion attacks
  Google's demonstration of extraction of training data that the GPT-2 model has memorized: overview / 🔗arXiv:2012.07805

• attacks on AI adversarial attacks poisoning attacks model inference attacks
  → Posts on PortSwigger's "The Daily Swig" by Ben Dickson