CDEF flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Consequentialism in data ethics

An ethical theory judging actions solely by outcomes, holding that the right act maximizes overall welfare. In data work it justifies harm-benefit analysis but can permit sacrificing minority interests for aggregate gain.

Deontological constraint

A duty or rule that binds regardless of consequences. Consent, honesty, and non-deception cannot be overridden merely because violating them would produce a better aggregate outcome for more people.

Veil of ignorance

A Rawlsian device for choosing principles without knowing one's own position in society. Applied to data systems, it asks whether you would accept the design if you might be the worst-off data subject.

Is-ought gap

Hume's observation that descriptive statements about what is the case cannot by themselves establish normative conclusions about what ought to be done. A value premise must be stated explicitly.

Data minimization principle

A data protection principle requiring that personal data be adequate, relevant, and limited to what is necessary for the stated purpose. Collecting extra data for hypothetical future use violates it.

Pseudonymization

Replacing identifiers with codes so data cannot be attributed to a person without additional separately held information. Pseudonymized data remain personal data under GDPR because re-identification remains possible.

Differential privacy epsilon

A parameter bounding how much a randomized query output can change when one individual's record is added or removed. Smaller epsilon means stronger privacy guarantees and typically more added noise.

k-anonymity limitation

k-anonymity ensures each record matches at least k-1 others on quasi-identifiers, but it fails to prevent attribute disclosure when all records in an equivalence class share the same sensitive value.

Statistical parity

A fairness criterion requiring equal positive decision rates across protected groups. It ignores differing base rates and can force decisions inconsistent with the underlying outcome distribution.

Equalized odds

A fairness criterion requiring equal true positive rates and equal false positive rates across groups. It conditions on the true outcome, so it accounts for base-rate differences that statistical parity ignores.

Fairness impossibility theorem

When group base rates differ, calibration within groups, balance for the positive class, and balance for the negative class cannot all hold simultaneously except in degenerate cases. A metric choice is therefore unavoidable.

Disparate impact rule

A screening heuristic under which a protected group's selection rate below eighty percent of the highest group's rate indicates potential adverse impact, prompting further justification or mitigation.

Proxy discrimination

Disparate impact transmitted through features correlated with a protected attribute, such as postcode or school name, even when the protected attribute itself is excluded from the model.

Post-hoc explanation

A method reconstructing reasons for an already trained model's output, such as LIME or SHAP. It approximates model behavior and must be validated for faithfulness and stability.

Right to explanation

The GDPR-derived expectation that data subjects receive meaningful information about the logic of automated decisions, their significance, and envisaged consequences, supporting contestation of outcomes.

Algorithmic impact assessment

A structured pre-deployment review covering purpose, data sources, affected stakeholders, risks, mitigations, and monitoring plans, repeated periodically or when the system's context changes materially.

Contestability

The ability of an affected person to challenge an automated decision and receive a reasoned response. It requires a channel, a decision-maker able to reverse the outcome, and a remedy.

Least privilege access

A governance control granting users only the data access needed for their current task, reviewed periodically. It limits breach blast radius and reduces incidental exposure of sensitive records.

Breach notification deadline

Under GDPR, controllers must notify the supervisory authority within seventy-two hours of becoming aware of a personal data breach unless it is unlikely to result in a risk to rights.

Data retention schedule

A governance artifact defining how long each data category is kept and how it is destroyed. It implements storage limitation and must extend to backups and derived datasets.

Controller versus processor

A controller determines the purposes and means of processing and bears primary accountability; a processor handles data on the controller's documented instructions and has narrower direct duties.

Data protection impact assessment

A GDPR Article 35 process required before high-risk processing such as large-scale monitoring or special-category data, documenting necessity, risks, and safeguards, with supervisory consultation if residual risk stays high.

GDPR maximum administrative fine

The higher of twenty million euros or four percent of total worldwide annual turnover, applicable to violations of core principles, legal bases, and data subject rights.

Cross-border transfer mechanism

A legal basis for moving personal data outside the EEA: an adequacy decision, standard contractual clauses, binding corporate rules, or a specific derogation, each requiring documented safeguards.

Precautionary principle

The rule that plausible but uncertain harm justifies proportionate preventive action before conclusive evidence arrives. It shifts the burden of proof toward demonstrating safety rather than waiting for proof of harm.

Reversibility test

A decision heuristic preferring options whose harms can be undone. Irreversible actions such as public data release or large-scale deployment require stronger evidence and broader sign-off.

Moral residue

The unmet obligation that remains after a justified ethical trade-off. Acknowledging and mitigating residue prevents a decision from being treated as fully resolved when some party still bears a cost.

Membership inference attack

An attack determining whether a specific individual's record was part of a model's training data. It is a privacy risk for generative and predictive models trained on personal data.

Federated learning privacy limit

Training on devices keeps raw data local, but shared model updates can leak training examples through gradient inspection, so secure aggregation and differential privacy are layered on.

Model drift

Degradation of a deployed model's performance as the real-world data distribution diverges from training data. Monitoring for drift is both an engineering and an ethical obligation.

Psychological safety

A team climate in which members can raise concerns and admit uncertainty without fear of punishment. It is a precondition for surfacing ethical risk before harm occurs.

Protected disclosure

A whistleblowing report made in good faith about a genuine legal or ethical breach through an internal or prescribed external channel, which many jurisdictions shield from retaliation.