Chapter 3 – Debugging toward Responsible AI
- Data collection bias: Data that is collected could have biases such as gender bias, as in Amazon’s applicant sorting examples, race bias, as in COMPAS, socioeconomic biases, as in hospitalization examples, or other kinds of biases.
Sampling bias: Another source of data bias could be in the process of sampling data points or sampling the population in the data collection stage of the life cycle. For example, in sampling students to fill in a survey, our sampling process could be biased toward girls or boys, rich or poor student families, or high- versus low-grade students.
- Perfect-knowledge white-box attacks: The attacker knows everything about the system.
Zero-knowledge black-box attacks: The attacker doesn’t have any knowledge of the system itself but collects information through predictions of the model in production.
- The encryption process transforms the information, data, or algorithm...
