Lecture 4 Flashcards
What forms can privacy take?
- privacy as confidentiality
- privacy as control
- privacy as normative ethical value (a snapshot)
Privacy as confidentiality
In classic information systems:
* access and sharing of data can be valuable (trust)
* sharing information can be dangerous (risks)
Confidentiality is one of the pillars of cybersecurity
* Keeping the information confidential is an obligation
* controlling access and usage is very important
* the ability to track and log access and usage is a legal obligation
platforms challenge our understanding of all the above
Privacy as control
information asymmetries, and power unbalances are present in many service models. This could lead to information totalirtarianism.
–> one party has all powers to collect and use data
According to the GDPR, a data subject should have rights and control over their data:
* informed consent and transparency
* right ot information about one’s data
* right to be forgotten
* accuracy
control through privacy-preserving technologies
Privacy as normative ethical value (A snapshot)
An ethical duty to safeguard our own and others’ privacy even when the use of our data is in out best interest
datafication implies there is a computable “truth” abiout the self. many argue that there should be parts of the self that not and should not be computable.
Our positionality is important to acknowledge when addresing an analysing (epistemology) questions related to privacy, power, fairness, justice etc.
What is positionality?
- the socail and political context that creates your identity and
- how your identity influences and biases your perception of and outlook on the world.
things like: race, ethnicity, gender, socioeconomic status, ability status, sexuality, age, Education, Experiences, Citizenship, Religious Beliefs, Marital Status, Education, Political Ideology, Appearance, Geographic, Location
Positionality statement
An acknowledgement (and awareness) of one’s own identity, social and political context, including privileges, experiences and emotions. And how these can influence our epistemology.
(in a positionality statement) objectivity claims have been challenged, especially in research including social components. Positionality is a form of transparency in situated subjectivity
That means that different persons and actors can and will have different understandings and positions.
- Positionality is malleable (easily influenced, pliable) and mutable (inconsistent, liable to change).
- It should be expressed contextually.
Statistical Thinking
“Statistical thinking is a way of understanding a complex world by describing it in
relatively simple terms that nonetheless capture essential aspects of its
structure.”
Building simplified models of complex phenomena to answer questions, find
patterns and correlations, validate hypotheses…
Often data-driven (quantitative)
Statistical thinking falls into 3 parts:
* Descriptive: I notice …
* inferential: I wonder …
* Contextual: I worry… & I expect …
Statistical thinking and bias:
“all models are … “
all models are Wrong, but some are useful
all models are biased, but some are useful
Some biases are useful, but biases can become harmful
Discrimination legal terminology
Discrimination can be due to
* Individual practices (I prefer men over women)
* Institutional practices (you cannot become prime minister unless born in Netherlands)
* structural practices
Dominant vs. Minority groups
* Dominant: Individuals or groups who have power in society
* Minority: groups that lack power often historically (should be protected)
Individual vs. Group discrimination
direct and indirect discrimination:
direct: formal or intentional discrimination of individuals or groups based on protected characteristics (e.g. age, disability, gender reassignment, race, sex, marriageand civil partnership)
indirect cases where an appearently neutral provision, criterion or practice would put persons of certain racial or ethnic origin (with a disability,
different sexual orientation, faith
and cultural practices etc.) at a particulare siadvantage compared with other persons.
Harms of discrimination
Allocative harms:
when gender, race and other protected charateristics play a role in the allocation of resources and opportunities.
Representational harms:
when a particular representation diminishes particular identities or population groups
What are protected characteristics (discrimination)?
– Age,
– Disability,
– Gender reassignment,
– Marriage and civil partnership,
– Pregnancy and maternity,
– Race,
– religion or belief,
– sex(gender), and sexual orientation.
Proxy of protected characteristic
A proxy is a seemingly “non-sensitive” or “non-protected” attribute that can be
used to infer a “sensitive” or “protected” attribute.
Inference can be of “logical” or “common sense” nature
* E.g., data about pregnancies can reveal that a person’s sex
Inference can be of a “statistical” nature: strong correlation in data
* E.g., in the US zip code often reveal the income and ethnicity
Or in other words:
proxy is something that is seemingly unharmful but this proxy attribute can become sensitive. e.g. from the postal code, you can understand the salary of the people or at least estimate it.
Bias in the ML pipeline
3 stages of bias processing:
* Pre-processing
* In-processing
* Post-processing
Historical Bias
Occurs when and ML model reproduces or
reinforces a harmful stereotypes already existing in the world.
for instance white defendants scores are skewed toward lower-risk categories (E.g. the higher the risk score, the lower the amount of people)
whereas for black defendantst the scores are not skewed toward lower-risk categories (e.g. all risk scores approximately the same amount of people).
Representation Bias
Representation Biases: occurs when the data is not representative of the population the model is being developed for, or specific categories
are largely under-represented.
Think of for instance recognizing software between men and women and black women are catergorized as men.
