Fairness, Privacy and Responsible AI
Measuring and reducing bias and protecting privacy in AI systems: harms and disaggregated evaluation, sources of bias and feedback loops, fairness definitions and impossibility results, mitigation, bias in generative models, re-identification and k-anonymity, differential privacy, federated learning and privacy-enhancing technologies, membership inference and unlearning, model cards and explainability, and audits, governance and monitoring.
AI-460 Year 4 3 credits 12 lessons · 4 h 58 min
Offered by HSM University · Artificial Intelligence
This is the Canadian version of this course. See the American version.
What you'll learn
- Explain the main kinds of AI harm, map responsible AI questions to each life cycle stage and run disaggregated evaluations
- Identify historical, representation, measurement, aggregation, learning, evaluation and deployment bias, including proxies and feedback loops
- Define and compute demographic parity, equalized odds, equal opportunity, predictive parity, calibration and individual and counterfactual fairness
- Explain the impossibility results linking calibration and error rates, apply the Chouldechova identity and reason about fairness and accuracy trade-offs
- Apply pre-processing, in-processing and post-processing mitigation, compare results and account for legal limits on disparate treatment
- Measure and mitigate bias in language and generative models with counterfactual tests, benchmarks and deployment safeguards
- Explain contextual integrity, quasi-identifiers and linkage attacks, and apply k-anonymity, l-diversity and privacy by design
- Define differential privacy, calibrate Laplace and Gaussian noise to sensitivity, and manage privacy budgets with composition
- Select privacy-enhancing technologies including DP-SGD, federated learning, secure aggregation, homomorphic encryption, TEEs and synthetic data for a threat model
- Run and interpret membership inference and extraction tests and explain defences and machine unlearning
- Write datasheets and model cards and produce global, local and counterfactual explanations while recognizing their limits
- Conduct impact assessments and audits, design governance structures and monitor fairness and privacy after deployment
Course content
12 lessons · 4 h 58 min total · final exam
- 1 What responsible AI means 25 min · quiz
- 2 Sources of bias in data and models 25 min · quiz
- 3 Defining and measuring fairness 25 min · quiz
- 4 Impossibility results and fairness trade-offs 24 min · quiz
- 5 Bias mitigation techniques 25 min · quiz
- 6 Fairness in language and generative models 25 min · quiz
- 7 Privacy foundations, re-identification and anonymization 25 min · quiz
- 8 Differential privacy 25 min · quiz
- 9 Privacy-preserving machine learning 25 min · quiz
- 10 Privacy attacks on models and machine unlearning 25 min · quiz
- 11 Documentation, transparency and explainability 25 min · quiz
- 12 Audits, governance and monitoring in practice 24 min · quiz
- E Final exam 20 questions
Requirements
Recommended background. You can start any course at any time.
Description
Measuring and reducing bias and protecting privacy in AI systems: harms and disaggregated evaluation, sources of bias and feedback loops, fairness definitions and impossibility results, mitigation, bias in generative models, re-identification and k-anonymity, differential privacy, federated learning and privacy-enhancing technologies, membership inference and unlearning, model cards and explainability, and audits, governance and monitoring.
Part of the Science stream, typically taken in year 4.