1. Introduction

What is the descriptive definition of Privacy?

Something that belongs to us and that is kept seperate from others.

• F.e. our homes, throughts, feelings, family life.

What is the normative definition of Privacy?

• Something that should not be accessed or known by others, certainly not the general public.

• We get to decide what happens with private things.

What is the definition of Privacy as the right to be let alone?

• By Warren and Brandeis (1890).

• Right to be let alone → freedom from intrusion by the press.

What is the definition of Privacy in Information and Communication?

The claim of individuals, groups or instistutions to determine when how and to what extent information about them is communicated to others.

What is the definition of Privacy in terms of Control?

Individuals ability to control the initial release of information about them and its subsequent distribution and use.

What is Microdata?

• Datasets containing information about specific individuals.

• Characterized by high dimensionality and sparsity.

What are Pricacy Risks of publishing Microdata?

• Removing identifying information (names, addresses, etc.) is not sufficient for privacy.

• Context / background knowledge / cross-correlation with public available data can be used to re-identifiy individual data records.

• Can reveal sensitive non-public information about a persons political or even sexual preferences.

What does Sparsity mean in Microdata?

A pair of random records are located far apart in the multi-dimensional space defined by the data attributes.

What is Collaborative Filtering?

Predicting a customers future choices using the knowledge of what similar customers did.

What are Linkage Attacks?

• Re-identify de-identified data.

• Works by identifying a digital fingerprint in the data → combination of features that uniquely identifies a person.

• Allows for linking a persons data from two different datasets.

What is a Membership Inference Attack on LLMs?

Infer wether an examples was in the training data.

What is a Data Extraction Attack on LLMs?

• Recover full training examples from LLM.

• An example x is Extractably Memorized if an adversary (without access to x) can construct a prompt that makes the model produce x.