A company initially intended to use a large data set containing personal information to train an Al model. After consideration, the company determined that it can derive enough value from the data set without any personal information and permanently obfuscated all personal data elements before training the model. This is an example of applying which privacy-enhancing technique (PET)?

The scenario describes the process of anonymization. The key action is that the company "permanently obfuscated all personal data elements." Anonymization is a data processing technique that irreversibly removes or modifies personally identifiable information (PII) from a dataset. The goal is to ensure that data subjects can no longer be identified, directly or indirectly. By doing this, the data ceases to be personal data, which aligns with the company's decision to derive value without processing personal information. This permanent and irreversible alteration is the defining characteristic of anonymization.

B. Pseudonymization: This is incorrect because pseudonymization is a reversible process. It replaces direct identifiers with pseudonyms, but the original data can be re-identified using additional information kept separately. The scenario specifies a "permanent" change.
C. Differential privacy: This is incorrect because it is a mathematical framework that adds statistical noise to the outputs of an analysis or model to protect individual privacy, rather than altering the source dataset by removing identifiers before training.
D. Federated learning: This is incorrect because it is a decentralized machine learning architecture where the model is trained on local data without the data leaving its source device. The scenario describes a data pre-processing step, not a training methodology.

1. National Institute of Standards and Technology (NIST). (2015). NISTIR 8053: De-Identification of Personal Information. U.S. Department of Commerce. In Section 2.1.2

the document defines anonymization as a process that removes the association between a set of identifying data and the data subject. It contrasts this with pseudonymization

where a link is maintained (Section 2.1.3). The "permanent" nature described in the question aligns with the goal of anonymization to sever this link. (Available at: https://csrc.nist.gov/publications/detail/nistir/8053/final)

2. European Union Agency for Cybersecurity (ENISA). (2018). Pseudonymisation techniques and best practices. In Section 2.1

"Basic Concepts

" the report clearly distinguishes between anonymisation and pseudonymisation

stating

"Anonymisation is the process of turning personal data into anonymous information... This process is irreversible." This directly supports why option A is correct and B is incorrect. (Available at: https://www.enisa.europa.eu/publications/pseudonymisation-techniques-and-best-practices)

3. Dwork

C.

& Roth

A. (2014). The Algorithmic Foundations of Differential Privacy. Foundations and Trends® in Theoretical Computer Science

9(3–4)

211–407. Chapter 2 defines differential privacy as a property of an algorithm's output

ensuring that the inclusion of any single individual's data in the input dataset has a statistically insignificant effect on the output. This confirms it is not a data removal technique as described in the question. (Available at: https://www.nowpublishers.com/article/Details/TCS-042)

4. Konečný

J.

McMahan

H. B.

Yu

F. X.

Richtárik

P.

Suresh

A. T.

& Bacon

D. (2016). Federated Learning: Strategies for Improving Communication Efficiency. arXiv:1610.05492. The paper's introduction (Section 1) defines federated learning as a technique where a shared model is trained across many clients

each with its own local data

without exchanging the data itself. This confirms it is a training architecture

not a data obfuscation method. (Available at: https://arxiv.org/abs/1610.05492)