DATA 120

Suppression

A data privacy technique that removes specific values, records, or columns from a dataset to reduce the risk of identifying individuals (e.g., deleting names or Social Security Numbers).

Generalization

A data privacy technique that replaces specific information with broader ranges or categories to make individuals harder to identify while preserving useful data (e.g., changing age 19 to age range 18–21).

K-anonymity

a privacy technique that ensures each person shares the same combination of identifying characteristics (like age, ZIP code, and gender) with at least k−1 other people, making them harder to identify.

Differential Privacy

A data privacy technique that adds controlled randomness (noise) to data analysis results to protect individuals' information while still allowing useful overall patterns to be identified (e.g., reporting 331 patients instead of the actual 327).

Example of Differential Privacy

Imagine a hospital wants to publish how many patients have diabetes. The hospital reports 331 instead of 327.

Because adding carefully controlled randomness makes it harder to determine whether any particular person's information was included.

Local Differential Privacy

a privacy technique where random noise is added to a person's data on their own device before it is shared, so organizations never receive the exact original information.

Examples of Local Differential Privacy

Imagine a school asks 10,000 students:

"Have you ever cheated on an exam?"

Here's how LDP works:

1. Student answers: A student selects "Yes."

2. Noise is added: Their phone randomly changes or keeps the answer using a privacy-protecting algorithm.

3. Data is sent: The school receives the randomized answer, not necessarily the student's real answer.

4. Results are analyzed: By combining thousands of randomized answers, the school can estimate how many students cheated without knowing any individual's true answer.

Federated Learning

a machine learning technique where multiple organizations or devices train a shared AI model without sharing their original data, keeping sensitive information stored locally.

Examples of federated learning

Imagine three hospitals want to train an AI model to predict which patients are at risk of developing a disease.

Normally, they might combine all their patient records into one central database. But that creates privacy risks.

Federated Learning solves this problem:

1. Keep data local: Each hospital keeps its patient records in its own database.

2. Train locally: Each hospital trains a copy of the AI model using its own patient data.

3. Share updates: Hospitals send model updates, not actual patient records, to a central server.

4. Combine updates: The server combines the updates to improve the shared model.

5. Repeat: The improved model is sent back to hospitals for further training.

Privacy Law: GDPR

Jurisdiction: EU

Who / What does it protect?: Personal data of individuals in the EU – broadly defined as information relating to an identified or identifiable person

When does it apply?: Applies to organizations established

in the EU that process personal data and can also apply to organizations outside the EU when they offer goods/services to people in the EU or monitor their behavior there

One important protection or rule: Individuals have significant rights over their personal data. Organizations must follow principles such as lawful processing, transparency, purpose, limitation, data minimization, security, and accountability.

Privacy Law: CCPA/CPRA

Jurisdiction: California

Who / What does it protect?: Personal information of California residents in their role as consumers subject to regulatory definitions and exceptions

When does it apply?: Applies to certain for-profit businesses doing business in California that meet statutory thresholds

One important protection or rule: Gives consumers rights such as knowing, deleting, correcting, and opting out of sale / sharing of personal information, plus protections involving sensitive personal information

Privacy Law: HIPAA

Jurisdiction: US Federal

Who / What does it protect?: Protected health information (PHI) handled by covered entities and their business associates

When does it apply?: Applies to covered healthcare providers, health plans, healthcare clearinghouses, and their business associates when handling PHI

One important protection or rule: Limits uses/disclosures of PHI and gives individuals certain rights to their health information

Privacy Law: FERPA

Jurisdiction: US Federal

Who / What does it protect?: Education records of students at educational institutions covered by FERPA

When does it apply?: Generally applies to educational agencies/institutions that receive funds under applicable US Dept of Education Programs

One important protection or rule: Gives parents and eligible students rights to access records and limits disclosures of PII from education records (subject to exceptions

Privacy Law: COPPA

Jurisdiction: US Federal

Who / What does it protect?: Personal Information collected online from children under 13

When does it apply?: Applies to operators of websites / online services directed to children under 13, and certain other operators with actual knowledge they are collecting personal information from children under 13

One important protection or rule: Requires notice and, in covered circumstances, verifiable parental consent before collecting personal information from children

AI life cycle

The AI Life Cycle is a step-by-step process used to plan, build, test, deploy, and continuously improve AI systems, while considering fairness, privacy, and ethical risks at every stage.

AI Lifecycle

Define Problem (Define the goal) → Collect Data → Measure → model → Evaluate → Deploy

Stereotyping

Groups are associated with stereotypical roles, traits, or behaviors

Assumptions

The model makes unsupported assumptions about identity or circumstances

Differences in treatment

Similar prompts receive different responses based on

demographic characteristics

Homogenization

Outputs lack diversity and repeatedly portray similar people, cultures, or perspectives

Societal impact

Outputs may reinforce stereotypes, exclusion, or existing

inequalities

Generative Ai Bias Audit

a systematic process used to test whether an AI model repeatedly produces unfair, stereotypical, or biased content, rather than just making a one-time mistake.

Advantages and Disadvantages of K-anonymity

Advantage: Relatiely intuitive; Measurable criterion; Preserves row-level data

Disadvantage: Can reduce datautility; difficult to achieve with many variables; difficult to choose correct quasi- identifiers; does not provide a universal privacy guarantee

Primary Privacy Risk that remain:Attribute Disclosure (characteristics may still be inferred); Background-knowledge / linkage attacks (outside information may reveal other information)

Advantages and Disadvantages of Differential Privacy

Advantage: Provides formal mathematical privacy

guarantee; protects against broad classes of information;

useful for aggregate staitistics and ML applications; allows

population info be learned without individual records

Disadvantage: Introduces noise; requires choices about how much error to allow; implementation can be technically difficult

Primary Privacy Risk that remain: Organization still posses the original data, so breaches, unauthorized access or misuse is still possible.

Advantages and Disadvantages of Local Differential Privacy

Advantage:Collector never receives the

exact value; reduces need to

trust central collector; useful

for collecting info from many

individual

Disadvantage: Usually requires more

noise than regular

differential privacy;

can require very large

samples; less suitable

for complex analyses

requiring precise

individual records

Primary Privacy Risk that remain: Device / software

implementation can

fail; metadata could

still identify users;

privatized reports still

require appropriate

handling

Advantages and Disadvantages of Federated Learning

Advantage: Avoids creating one giant

centralized raw-data

repository; enables

collaboration without data

sharing; raw data remain

under local control

Disadvantage: Complex

infrastructure;

communication /

computation costs;

heterogeneous data

can affect model

quality; not inherently

a formal privacy

guarantee

Primary Privacy Risk that remain: Model updates can

leak information

about training

records; local

databases can still be

breached; final

models can

sometimes leak

training information

What is GDPR and who does it protect?

EU privacy law protecting personal data of identifiable individuals

When does GDPR apply and what are its key rules?

Applies to EU organizations and certain non-EU organizations serving or monitoring people in the EU. Requires lawful processing, transparency, data minimization, and individual rights

What is CCPA/CPRA and who does it protect?

California privacy laws protecting California residents' personal information.

When does CCPA/CPRA apply and what rights does it give?

Applies to certain for-profit businesses meeting legal thresholds. Rights include knowing, deleting, correcting, and opting out of data sale/sharing.

What is HIPAA and who does it protect?

US federal law protecting health information (PHI) held by covered healthcare entities and business associates.

When does HIPAA apply and what is its key rule?

Applies to covered healthcare providers, health plans, clearinghouses, and business associates. Limits PHI use and disclosure.

What is FERPA and who does it protect?

US federal law protecting student education records at covered educational institutions.

When does FERPA apply and what is its key rule?

Generally applies to schools receiving applicable federal education funding. Gives access rights and limits disclosure of student records.

What is COPPA and who does it protect?

US federal law protecting online personal information collected from children under 13.

When does COPPA apply and what is its key rule?

Applies to child-directed online services and certain services knowingly collecting data from children under 13. Generally requires parental notice and verifiable consent.