A medical record contains a name.
Remove the name.
Is the person now anonymous?
Not necessarily.
Age, postcode, rare diagnosis, admission date, occupation and unusual combinations of facts may still identify someone when linked with other information.
Privacy is therefore not a single delete key.
Quick Read
Anonymisation and de-identification transform data to reduce the risk that records can be linked back to individuals or reveal sensitive attributes. Techniques include removing direct identifiers, generalising values, suppressing rare records, perturbing data, aggregating groups and, in some settings, applying formal privacy mechanisms such as differential privacy.
The central tension is unavoidable: stronger privacy protection can reduce analytical utility because relationships, granularity and rare cases may be altered or removed. Contemporary privacy research therefore treats anonymisation as a privacy–utility trade-off rather than a magical state in which data remain fully useful while identity risk disappears.
One-sentence answer: anonymisation is lossy because protecting people often requires deliberately reducing the precision, linkability or uniqueness of the data that described them.
Names Are the Easy Part
Direct identifiers such as names, identity numbers, email addresses and phone numbers are obvious.
Remove them and a dataset is safer.
But identity can survive in combinations.
A 97-year-old violin maker living in one small district may be recognisable without a name. A rare disease combined with a public news story can create an anchor. A precise sequence of locations can reveal a home and workplace.
Anonymisation therefore has to reason about uniqueness, linkage and adversarial background knowledge.
Quasi-Identifiers: Ordinary Facts That Become Identifying in Combination
Age is not usually a name.
Postcode is not usually a name.
Sex, occupation and date are not names.
But combined, they can narrow the candidate set dramatically.
These quasi-identifiers reveal why privacy risk is relational. A field can be harmless alone and identifying when connected to external data.
Suppression: Remove the Risky Value
One privacy strategy is suppression.
A rare occupation becomes blank. An unusually specific location disappears. An entire outlier record may be removed.
This reduces re-identification risk because the distinguishing feature is no longer available.
It is also direct information loss.
If rare cases are scientifically important, the data may become safest precisely where it becomes least informative.
Generalisation: Make the Person Less Unique by Making the Data Coarser
Age 47 becomes age 45–49.
A six-digit postcode becomes a broader district.
An exact date becomes a month or year.
Generalisation reduces resolution so more individuals share the same representation.
The privacy gain comes directly from lossy aggregation of identity-relevant detail.
k-Anonymity: Hide Inside a Crowd
One classic privacy idea is k-anonymity.
Informally, a released record should be indistinguishable from at least k−1 others with respect to selected quasi-identifiers.
This can be achieved through generalisation and suppression.
It reduces one form of re-identification risk, but it does not solve every privacy problem. If everyone in an equivalence class shares the same sensitive diagnosis, learning the class can still reveal the diagnosis.
Privacy is more than hiding a row among similar rows.
l-Diversity and t-Closeness Tried to Protect Sensitive Attributes
Extensions such as l-diversity and t-closeness were developed to address weaknesses in k-anonymity by considering the diversity or distribution of sensitive attributes within groups.
Each refinement reveals the same pattern: privacy risk depends not just on identifiers but on what can be inferred after the transformation.
Differential Privacy Changes the Question
Differential privacy does not try to make every individual record look like several others.
Instead, it provides a formal guarantee limiting how much the presence or absence of one person can change the distribution of released outputs under a specified privacy budget.
Noise is deliberately introduced or queries are otherwise controlled so aggregate information can be released with bounded individual influence.
Again there is a bargain: stronger privacy generally requires more distortion or less query freedom.
The Privacy Budget Is a Distortion Budget in Reverse
In ordinary compression, we ask how much distortion can be tolerated to save bits.
In privacy-preserving release, we ask how much utility can be preserved while limiting what an observer can learn about an individual.
The optimisation target changes, but the architecture feels familiar: preserve enough structure for the intended receiver while deliberately destroying structure that creates unacceptable risk.
Privacy and Utility Do Not Move Independently
Recent clinical-data research makes this concrete.
Studies comparing de-identification configurations find that reducing re-identification risk can suppress records, mask variables and change distributions enough to degrade downstream predictive utility.
This does not mean privacy protection is bad.
It means privacy transformations must be evaluated against the actual analytical job rather than scored in isolation.
Rare Cases Are Where Privacy and Science Collide Most Sharply
Common patterns are easy to hide inside groups.
Rare combinations are difficult.
Yet rare combinations are often exactly what medical research, safety engineering and anomaly detection care about.
A privacy method can therefore erase the tail of a distribution—the same tail where consequential discoveries live.
Location Data Is Especially Difficult
A sequence of precise locations can identify routines.
Home at night. Workplace in the day. A clinic on Tuesday. A school pickup at 3 pm.
Removing a name does little if the trajectory itself is distinctive.
Protecting location often means reducing spatial resolution, temporal resolution or both, which directly reduces what mobility analysis can recover.
Speech and Faces Show That Identity Lives in the Signal
Voice is not merely the words spoken.
Pitch, timbre, accent, rhythm and learned characteristics can support identity inference.
Face imagery carries identity in geometry, texture and context.
Anonymising such data may preserve content while altering identity-bearing features. Recent speech-anonymisation research explicitly studies the balance between privacy and retained clinical or research utility.
Synthetic Data Is Not Automatically Anonymous
Generating synthetic records can reduce direct exposure of real records.
But synthetic data can still leak information if a model memorises training examples or preserves rare combinations too faithfully.
Recent studies therefore evaluate synthetic data along three separate axes: fidelity, utility and privacy.
A dataset can score well on one and poorly on another.
Anonymisation Can Change Correlations
Suppose age is generalised into broad bands and several rare occupations are suppressed.
Relationships involving age or occupation may weaken, flatten or change shape.
A model trained on the anonymised dataset is therefore not necessarily modelling the same statistical landscape as one trained on the original.
Privacy transformations are part of the data-generating pipeline from the analyst’s point of view.
Re-Identification Risk Changes When External Data Changes
A dataset considered difficult to re-identify today may become easier tomorrow if new public datasets appear.
Privacy risk is therefore not entirely intrinsic to the released table. It depends on the surrounding information environment and plausible attacker capabilities.
Anonymity can age.
Education: Anonymous Data Can Protect Students and Flatten Them
Schools may want to analyse learning patterns without exposing individual students.
Aggregating results by cohort can protect privacy and reveal broad trends.
But aggressive aggregation can hide small groups, unusual learning needs or local failures.
The correct resolution depends on the decision and the duty to protect the child.
When Anonymisation Is Exactly the Right Loss
Some distinctions should not survive into every downstream representation.
A public statistical report may need age bands, not birthdays. A research dataset may need regional mobility patterns, not home addresses. A classroom case study may need learning characteristics, not a child’s name.
Loss can be ethical design.
When Anonymisation Becomes Dangerous
- Removing names is mistaken for full anonymity.
- Rare combinations remain re-identifiable.
- Generalisation destroys the subgroup needed for analysis.
- Privacy scores are optimised without testing downstream utility.
- External linkage data makes yesterday’s safe release risky.
- A synthetic dataset is assumed private without attack-based evaluation.
- Protected characteristics survive through proxies even when explicit columns are removed.
A Practical Anonymisation Audit
- Threat: what re-identification or inference attack is being protected against?
- Identifiers: which direct and quasi-identifiers remain?
- Transformation: what was suppressed, generalised, perturbed or aggregated?
- Utility: which analytical relationships must survive?
- Rare cases: what happens to outliers and minority groups?
- External data: what other datasets could be linked?
- Time: will privacy risk change as new data becomes public?
- Validation: has the released representation been tested for both privacy and task utility?
Sources and Further Reading
- 2024 clinical study of privacy–utility trade-offs under de-identification.
- 2025 research on privacy and utility in speech anonymisation.
- 2025 study of fidelity, privacy and utility in synthetic patient data.
- Foundational literature on k-anonymity, l-diversity, t-closeness and differential privacy.
Continue Through eduKateSG
Continue with How Lossy Works | Aggregation, How Lossy Works | Resolution, and How Lossy Works | Feature Selection.
Final Thought: Privacy Works by Making Some Questions Harder to Answer
That is not a bug.
It is the point.
The art of anonymisation is deciding which questions must become harder so people become safer, while preserving enough structure for legitimate knowledge to remain possible.