Ethics, Security and Impact

Objective: integrate risks into the life cycle, not after deployment.

Bias and fairness

Data reflect collection, access, history and human decisions. Measure errors and selection rates by relevant groups, but do not choose a fairness metric without legal and business context. Several fairness criteria may be mathematically incompatible when base rates differ.

Privacy

Minimize data, control access, encrypt, define retention and deletion. A model can memorize rare examples. Test exposure and avoid publishing weights trained on sensitive information without analysis.

Threats

• poisoning of training data;
• adversarial examples and out-of-distribution inputs;
• model extraction by queries;
• inversion or membership inference;
• untrusted dependencies, models and weights files.