Eye Information Privacy
Research Output:
Theses
PhD thesis
Research
Abstract
This thesis examines the privacy of eye information, including eye images, gaze, eye movements, and related eye tracking data and the information they can convey about individuals. It asks when eye information becomes privacy relevant, what existing safeguards actually protect, how claims of protection can be evaluated, and how protection should respond when processing, interpretation, sharing, and reuse change what eye information reveals and the conditions under which it is used.<br/>The thesis first establishes the conceptual, methodological, legal, and technical foundations needed to examine these questions (Chapters 2 and 3), before evaluating how privacy has been defined, studied, and protected across the existing eye tracking literature (Chapter 4). Two complementary reviews of a corpus of 78 publications combine ensemble non-negative matrix factorization, BERTopic, and a manually developed taxonomy of privacy harms to map recurring technical and thematic structures and examine how privacy risks and protections are framed. The reviews show that eye tracking privacy research is fragmented across data types, mechanisms, processing stages, and application contexts. Privacy claims often differ in the information being protected, the harms and actors considered, the assumptions under which protection is evaluated, the utility that must be retained, and the evidence used to support protection claims. These differences make safeguards difficult to compare and can leave unclear what risks remain after protection is applied.<br/>The reviews further show that protection at one representation or processing stage does not determine what later analysis may reveal, raising the question of whether existing legal classifications remain adequate when identifiable eye information supports probabilistic health inference (Chapter 5). In a clinical study of persistent post concussion symptoms, behavioral, gaze, and oculomotor measures supported individual level health related inference across five eye tracking tasks, although predictive uncertainty and class overlap limited standalone diagnosis. The findings distinguish clinical sufficiency from privacy and legal significance, showing that uncertain inference can still reveal health related information and that legal classification must account for what later processing can reveal, not only how the data were originally collected or described.<br/>Downstream reuse raises the related question of whether later use can remain observable after eye information has been released and direct control has weakened (Chapter 6). Radioactive watermarking is evaluated as a traceability mechanism for eye image datasets, providing statistical evidence that marked images contributed to later model training while largely preserving image quality and gaze classification performance. The method preserves evidence that can support scrutiny of whether subsequent use is consistent with the conditions under which the information was provided. The findings show that downstream protection may therefore depend not only on restricting access or use, but also on preserving the ability to observe and assess later use.<br/>These findings show that legal classification, technical safeguards, and governance are each necessary components of protection, but none alone determines whether an individual is adequately protected. Privacy of eye information should therefore not be considered established merely because a particular transformed representation is treated as privacy protected or because a single safeguard has been shown to be effective. It depends on whether the individual remains protected as eye information is transformed, acquires new meaning, and is used for new purposes.
Publication Information
Output type
Research Output:
Theses
PhD thesis
Research
Original language
EnglishPublication milestones
- Published - 2026
Publication status
Published - 2026
Supervisors/Advisors
- Dan Witzner Hansen (Principal Supervisor)
Access to documents
Submitted manuscript, 25.03 MB
