Privacy for AI and Analytics

By the end

What you'll build

• Trace the provenance of a training or evaluation dataset and judge whether the intended use is compatible
• Apply minimisation, pseudonymisation and aggregation choices inside an analytics pipeline
• Assess re-identification, memorisation and inference risk in a model or its outputs
• Design meaningful human oversight for a decision supported by a model
• Draft the privacy sections of an assessment for an AI or analytics use case
• Explain to a non-technical stakeholder what a model does with personal data and why

Curriculum

What's inside

8 modules · 51 lessons

1. 01

   Where the data came from

   6 lessons
   • Provenance and compatible use: whether this dataset may lawfully train that model
   • Secondary use and the compatibility assessment
   • Scraped, purchased and partner-supplied data
   • Consent, notice and the data people never expected to be modelled
   • Documenting provenance so it survives staff turnover
   • + 1 more lesson
2. 02

   Minimisation inside a pipeline

   7 lessons
   • Cutting fields without breaking the model: minimisation as an engineering choice
   • Feature selection as a privacy decision
   • Pseudonymisation, tokenisation and key management
   • Aggregation, sampling and synthetic alternatives
   • Differential privacy and noise, explained without the maths
   • + 2 more lessons
3. 03

   Inference and re-identification

   7 lessons
   • What a model can reveal that you never put into it
   • Linkage attacks and the uniqueness problem
   • Memorisation and training-data extraction
   • Membership inference in plain terms
   • Proxy variables and sensitive attributes by the back door
   • + 2 more lessons
4. 04

   Transparency and explanation

   6 lessons
   • Explaining an automated decision to the person it affected
   • Model documentation and data sheets
   • Notice for analytics and profiling
   • Meaningful information about logic, in practice
   • Explaining limits and uncertainty honestly
   • + 1 more lesson
5. 05

   Human oversight and contestability

   6 lessons
   • Designing oversight that is real rather than a rubber stamp
   • Where in the workflow a human can actually intervene
   • Reviewer skill, workload and automation bias
   • Appeal and correction routes for affected people
   • Logging decisions for later review
   • + 1 more lesson
6. 06

   Assessing the use case

   7 lessons
   • Running a privacy impact assessment on a model instead of a database
   • Framing the processing when it is probabilistic
   • Risk to individuals versus risk to the organisation
   • Mitigations that actually reduce risk
   • When to recommend not proceeding
   • + 2 more lessons
7. 07

   Operating and retiring the system

   7 lessons
   • Deletion, drift and retention when the data is inside the weights
   • Honouring erasure in a modelled world
   • Retention of prompts, logs and embeddings
   • Monitoring for drift with privacy-safe metrics
   • Hosted-model and third-party processing arrangements
   • + 2 more lessons
8. 08

   Practice and check

   5 lessons
   • Terms that get misused in model reviewsmatch pairs
   • Clearing a model before it touches productionsequence order
   • Getting the claim rightfill blank
   • The churn model that read the transcriptsscenario
   • Course quizquiz

The shape of it

How this course works