Machine learning is a data driven process and, in most cases, the algorithm’s performance improves when using more data. This characteristic of machine learning algorithms is a central drive for collecting and processing more data. Although there are many benefits that can be harvested from processing personal or sensitive data, it is essential to consider and assess the potential privacy violations in using this data.
Moreover, if the machine learning model you are developing makes decisions about individuals that can have a negative impact on their life – e.g. an automatic loan system that may automatically refuse loans – it is crucial to assess the algorithm’s fairness and inclusiveness for all members of the society. Moreover, it is crucial to assess that decisions are based on clear and interpretable features.
Whenever processing personal information or whenever developing algorithms that take automated decisions, consider to:
- assess that personal data used does not breach privacy,
- use privacy preserving machine learning whenever possible, e.g. differential privacy,
- define and implement metrics for bias, fairness and responsible use of machine learning,
- take security into account – confidential information can be leaked by machine learning algorithms,
- continuously monitor that the algorithm behaves responsibly,
- be as transparent as possible about the data used and the model you are developing.
- Test for Social Bias in Training Data
- Assess and Manage Subgroup Bias
- Use Privacy-Preserving Machine Learning Techniques
- Share Status and Outcomes of Experiments Within the Team
- Continuously Monitor the Behaviour of Deployed Models