Getting Started with AI Hacking: Part 1

This blog post introduces ‘AI hacking’ with a focus on attacks against classifier models, covering adversarial examples (image perturbations and physical-world attacks), techniques to evade ML-based antivirus/EDR (including shifting code attributes), model extraction (cloning), model inversion (training-data extraction), and data poisoning. It illustrates concepts with examples and research outcomes and outlines mitigations such as adversarial training, rate limiting, differential privacy, and data vetting.