Cloud Security Podcast by Google EP68 How We Attack AI? Learn More at Our RSA Panel!
6 snips
Jun 6, 2022 AI Snips
Chapters
Transcript
Episode notes
AI Threat Models
- AI threat models must consider new threats like data poisoning, where malicious data influences algorithms.
- Unlike traditional code, AI algorithms learn from external data, making them vulnerable to manipulation.
Adversarial Perturbations
- AI models are robust to random noise but vulnerable to adversarial perturbations.
- Small, crafted changes to input data can cause misclassification, even if the model is accurate in typical scenarios.
AI vs. Data Processing Systems
- Attacking AI differs from attacking complex data processing systems due to the unique nature of AI's threat model.
- AI's high-dimensional data and query capabilities make it susceptible to targeted attacks and extraction.