PaperlayerView exact revision on arXiv

Source-linked AI summary

Privacy-Aware Eye Tracking Using Differential Privacy

Julian Steil, Inken Hagestedt, Michael Xuelin Huang, Andreas Bulling

arXiv:1812.08000v3cs.HC

TL;DR

Eye movements can reveal sensitive information, while privacy-aware eye tracking remains under-explored. The paper pairs a survey of sharing concerns with differential privacy for a VR gaze interface, and reports protection of private information while retaining utility-task performance.

  • Problem

    Eye movements can reveal sensitive information, but privacy-aware eye tracking and evidence about users’ concerns remain under-explored.

    Source: §1 · p. 1, Privacy-Aware Eye Tracking Using… · p. 10

  • Method

    The paper surveys eye-tracking privacy concerns and applies differential privacy to gaze data in a VR interface.

    Source: §1 · p. 2

  • Results

    On a 20-participant dataset, the evaluation demonstrated protection of private information while maintaining performance on the document-type utility task.

    Source: §6.2 · p. 7, §8 · p. 8, §6.3 · p. 8

  • Takeaways & Limitations

    The authors conclude that differential privacy can preserve private information while maintaining utility in their gaze-interface evaluation.

    Source: §8 · p. 8

    The example study focuses on reading and was evaluated on a typical HCI dataset size; the authors expect generalization but note the privacy-utility trade-off may vary with sensitivity and data-vector length.

    Source: §7.2 · p. 8

Abstract

from arXiv · show

With eye tracking being increasingly integrated into virtual and augmented reality (VR/AR) head-mounted displays, preserving users' privacy is an ever more important, yet under-explored, topic in the eye tracking community. We report a large-scale online survey (N=124) on privacy aspects of eye tracking that provides the first comprehensive account of with whom, for which services, and to what extent users are willing to share their gaze data. Using these insights, we design a privacy-aware VR interface that uses differential privacy, which we evaluate on a new 20-participant dataset for two privacy sensitive tasks: We show that our method can prevent user re-identification and protect gender information while maintaining high performance for gaze-based document type classification. Our results highlight the privacy challenges particular to gaze data and demonstrate that differential privacy is a potential means to address them. Thus, this paper lays important foundations for future research on privacy-aware gaze interfaces.

1 INTRODUCTION

Eye movements can reveal sensitive information, yet privacy-aware eye tracking remains under-investigated. The paper addresses gaps in understanding users’ concerns and in privacy-preserving methods through a survey and a differential-privacy gaze interface.

1 INTRODUCTION

  • The eyes are typically not consciously controlled, can reveal unique private information, and are difficult to reconstruct in retrospect.These properties make it harder for users to learn from past privacy-related behavior.Source: §1 · p. 1
  • Privacy-aware eye tracking remains under-investigated, with gaps in both evidence about user concerns and methods, systems, and interfaces that protect privacy.The paper frames its work as addressing both gaps.Source: §1 · pp. 1–2
  • The paper combines a survey of eye-tracking privacy concerns with a differential-privacy method and a sample VR gaze interface.Differential privacy adds noise to reduce inference of sensitive information while retaining data utility for tasks such as document-type classification.Source: §1 · p. 2

2.1 Information Available in Eye Movements

  • Eye movements reveal varied information, including cognitive states, activity, personality traits, gender, and age, while wider access to eye tracking poses privacy risks.Prior work also links pupil size to interest and cognitive load, and eye movements to several mental disorders.Source: §2.1 · p. 2

2.2 Eye Movements as a Biometric

  • Prior work developed eye-movement biometrics for authentication, but the authors identify their study as the first practical exploration of eye tracking as a privacy threat.Earlier biometric work included task-independent authentication and user identification across everyday tasks.Source: §2.2 · p. 2

2.3 Differential Privacy

  • Differential privacy has been applied to other data types, but the authors report being the first to apply it to eye-tracking data.Its central challenge is balancing privacy against data utility, with mechanisms tailored to a use case offering a possible trade-off.Source: §2.3 · p. 2

3 PRIVACY CONCERNS IN EYE TRACKING

The survey finds that willingness to share gaze data depends on the service, inferable attributes, data recipient, setting, and use. Participants were more willing to share aggregated features and anonymized data.

Survey Method and Participants

  • The online survey used 7-point Likert questions and included 124 participants, predominantly young, educated people with technical backgrounds; the authors note likely self-selection among people interested in the topic.Recruitment used social platforms and local mailing lists; responses were grouped as disagree, neutral, or agree for analysis.Source: §3 · pp. 2–3

Survey Analysis

  • Nearly all answers differed significantly from an equal distribution under Pearson’s chi-squared test (p < 0.001, dof = 6).The detailed numerical results and plots are reported in supplementary material.Source: §3 · p. 3

Services and Attributes

  • More than 80% agreed to share data for disease detection; willingness varied by service, and a majority opposed sharing when sexual preference, gender, race, or identity could be inferred.A clear majority opposed sharing for shopping assistance and interest detection, while service responses for improved gaze-target recognition, website content, and activity recognition showed two peaks.Source: §3 · p. 3, Figure 2 · p. 3

Whom and Where

  • Opinions on sharing in general were split; participants favored government health-agency co-ownership and research, rejected company ownership, and accepted constrained environments and VR/AR collection over benefits.Participants opposed sharing in public or private environments and sharing in exchange for benefits.Source: §3 · p. 3, Figure 3 · p. 3

Data Representation

  • Participants largely preferred sharing aggregated statistical features, while gaze-revealing representations such as fixations and scan paths would mostly not be shared.Willingness to share increased when data was anonymized.Source: §3 · p. 4

4 PRIVACY-PRESERVING EYE TRACKING

This section frames privacy-preserving eye tracking as balancing formal privacy guarantees with useful gaze-based tasks. It defines two threat models and describes a differential-privacy mechanism applied before data release.

4 PRIVACY-PRESERVING EYE TRACKING

  • The key challenge is protecting privacy without letting performance on the desired eye-tracking task drop too far.The proposed mechanism is applied by a trusted curator before releasing aggregated gaze features.Source: §4 · p. 4

4.1 Threat Models

  • Without prior knowledge, an attacker tries to infer a private attribute such as gender; with prior knowledge, the attacker tries to re-identify a target.The second scenario assumes the attacker can link the target’s work data to their identity and then identify their leisure data.Source: §4.1 · p. 4

4.2 Differential Privacy for Eye Tracking

  • Differential privacy aims to make a mechanism’s answer independent of whether one person contributed data, typically by adding random noise.The section illustrates this with an average fixation-rate query that should be almost unchanged when one person’s data is included or excluded.Source: §4.2 · p. 4
  • The exponential mechanism favors higher-utility outputs, while L1 sensitivity bounds how much one participant can change the answer and sets the required noise scale.The paper describes sensitivity as the worst-case influence of one person’s data, including when that person is an outlier.Source: §4.2 · p. 5

4.3 Implementing Differential Privacy

  • The feature mechanism uses participant feature vectors and accounts for varying lengths by padding shorter vectors to a shared maximum length.Sensitivity is derived from feature ranges, which may be estimated from data or given by theoretical constraints.Source: §4.3 · p. 5
  • Because feature ranges differ, the implementation derives a separate mechanism for each feature and combines them using the composition theorem.It uses an L1-distance utility function for the derivation.Source: §4.3 · p. 5, §4.2 · p. 5
  • Subsampling one value per nonoverlapping window for each participant-feature further decreases sensitivity by a factor of the window size.The subsampling windows are independent of the feature-generation process.Source: §4.3 · p. 5

5 DATA COLLECTION

The study recorded eye movements while participants read three document types in VR, creating a dataset for document-type analysis. The protocol and extracted features were designed to capture reading behavior across those sessions.

5 DATA COLLECTION

  • Each participant read a comic, an online newspaper, and a textbook, with each document taking about 10 minutes to read.The documents varied in their proportions of text and images.Source: §5 · p. 6, Figure 5 · p. 6
  • The dataset includes 20 participants, recruited through university channels, who read in an Oculus DK2 headset fitted with a Pupil eye-tracking add-on.Participants were 21–45 years old, with 10 male and 10 female participants.Source: §5 · p. 6
  • Participants read in arbitrary order without eye-tracker calibration, allowing the researchers to record natural eye-movement behavior.After each document, they answered questions about fatigue, liking, understanding, difficulty, and text comprehension.Source: §5 · p. 6
  • The researchers extracted 52 features covering fixations, saccades, blinks, and pupil diameter using 30-second sliding windows stepped every 0.5 seconds.They also computed wordbook features encoding sequences of n saccades.Source: §5 · p. 6

6 EVALUATION

The evaluation tests gaze-based document classification as a utility task and gender prediction and user re-identification as privacy attacks. Across the tested noise settings, the reported results show reduced attacker performance alongside retained document-classification accuracy.

6.1 Classifier Training

  • The evaluation used SVM classifiers, with leave-one-person-out testing for document type and gender, and a train/test split for the prior-knowledge re-identification threat model.Document classification used 19 participants for training and one for testing; re-identification used the first and remaining halves of aggregated feature vectors, respectively.Source: §6.1 · pp. 6–7, §6 · p. 6
  • The DP mechanism applied the exponential mechanism across 20 participants and 52 features, using windows of 10 samples and feature ranges to set sensitivity.Experiments were repeated five times and results averaged to account for subsampling and noise generation.Source: §6.1 · p. 7

6.2 Without Prior Knowledge

  • At ϵ = 15, gender prediction reached the 54% chance level, and the authors conclude gender privacy is preserved on this dataset for ϵ ≤ 15.The chance level reflects slightly imbalanced classes in leave-one-person-out evaluation.Source: §6.2 · p. 7, Figure 6 · p. 7
  • At ϵ = 15, document-type classification accuracy was about 55%, approximately 22% above chance, despite training and testing on noised data.The authors report that noise affected this utility task less than gender prediction.Source: §6.2 · p. 7, Figure 6 · p. 7
  • With clean-trained classifiers, ϵ = 20 put gender prediction below chance while document classification remained around 70%; ϵ = 30 also preserved privacy.The authors report this setting as relevant to current datasets, which are not noised.Source: §6.2 · p. 7, Figure 7 · p. 7

6.3 With Prior Knowledge

  • With prior knowledge trained on clean data, at ϵ = 15, test-data noise reduced attacker accuracy to near chance (6.4%), while document-type utility retained about 70% accuracy.Source: Figure 8 · p. 7, §6.3 · p. 8

7 DISCUSSION

The discussion reports survey findings about users’ willingness to share gaze data and describes a differential-privacy VR method evaluated for privacy protection alongside gaze-based document classification.

7.1 Privacy Concerns in Eye Tracking

  • Survey participants were willing to share gaze data for medical applications or improved VR/AR experiences, but refused interest-identification and shopping-assistance services.A majority disliked services inferring identity, gender, sexual preference, or race; the authors suggest participants may not fully understand what gaze data can reveal.Source: §7.1 · p. 8
  • Adding noise to raw eye-movement data may not protect private attributes because smoothing can remove it; the authors recommend statistical or aggregated features instead.These features summarize temporal and appearance statistics across fixations, saccades, and blinks.Source: §7.1 · p. 8

7.2 Privacy-Preserving Eye Tracking

  • The VR evaluation used a new 20-participant dataset to study threat models with and without prior target-user knowledge and training with or without clean data.It evaluated gender inference and user identification as privacy-infringing tasks.Source: §7.2 · p. 8
  • Privacy was preserved with acceptable utility-task accuracy across evaluation paradigms for gaze-based document classification protecting users’ gender and identity.The example study tested whether differential privacy could protect those attributes while retaining document-type classification utility.Source: §7.2 · p. 8
  • The study focused on reading, and the privacy parameter ϵ may vary with sensitivity and data-vector length.The authors expect generalization to other eye-tracking activities and larger datasets, but describe these as expectations.Source: §7.2 · p. 8

8 CONCLUSION

The paper reports a large-scale survey of eye-tracking privacy concerns and presents a differential-privacy gaze interface evaluated in VR. On a new 20-participant dataset, the method preserved private information while maintaining utility-task performance.

8 CONCLUSION

  • A new 20-participant dataset evaluation demonstrated privacy protection while maintaining performance on a utility task.The paper frames this as implementing privacy without impeding utility.Source: §8 · p. 8
  • The paper reports the first large-scale online survey to understand privacy concerns about eye tracking and eye-movement analysis.The survey findings motivated the privacy-aware gaze interface.Source: §8 · p. 8
  • The authors chose a VR gaze interface in view of the significant and imminent threat potential from upcoming eye-tracking-equipped VR headsets.Source: §8 · p. 8

Privacy-Aware Eye Tracking Using Differential Privacy [Supplementary Material]

The paper presents a survey of gaze-data sharing preferences and a differential-privacy VR interface. Its evaluation reports protection of user identity and gender while retaining high performance on document-type classification.

Privacy-Aware Eye Tracking Using Differential Privacy

1 SURVEY RESULTS

The survey asked about eye-tracking privacy concerns, future uses, data sharing, and willingness to share different eye-movement representations. Responses used a seven-point Likert scale, with scores grouped for analysis.

1.1 Services and Attributes

  • The services-and-attributes questions asked about willingness to share eye-tracking data with services and agreement with those services inferring private attributes.Source: §1.1 · p. 11

1.2 Whom and Where

  • The whom-and-where questions covered general sharing, data recipients, environments, and sharing in exchange for benefits or VR/AR use.Source: §1.2 · p. 11

1.3 Data Representation

  • The data-representation questions compared willingness to share unmodified data with willingness to share anonymized data.Source: §1.3 · p. 11

2 STATISTICAL TESTS

Nearly all answers differed significantly from an equal distribution, and the majority of questions showed a significant difference from their corresponding normal distribution. The section also names the topics represented in Figures 1–3.

2 STATISTICAL TESTS

  • Nearly all answers differed significantly from an equal distribution (Pearson’s chi-squared, p < 0.001, dof = 6).Source: §2 · p. 12
  • Skewness analysis found that the majority of questions differed significantly from their corresponding normal distribution (p < 0.1).Source: §2 · p. 12

2.1 Statistical tests corresponding to Figure 1

  • Figure 1’s labels include services, private attributes, visual search, and target detection.Source: §2.1 · p. 12

2.2 Statistical tests corresponding to Figure 2

  • Figure 2’s labels include recognized private company, sharing owner environment application, and governmental health authority.Source: §2.2 · p. 12

2.3 Statistical tests corresponding to Figure 3

  • Figure 3’s labels include no modification, modified representation (anonymised), and features recognized private company.Source: §2.3 · p. 12

3 SURVEY EVALUATION

This section describes survey questions on participants’ familiarity with eye tracking and VR/AR, potential uses of gaze data, and willingness to share it under different contexts and collection conditions. It covers sharing in different representations—including behavioral measures, scene video, and eye-camera video—and hypothetical sharing after anonymization; the supplied passages report questions and response options, not respondent findings.

3.1 Eye Tracking and Virtual Reality (VR) Technologies

  • The survey asked about familiarity with eye tracking and VR/AR and concerns involving social, mental, physical, and privacy impacts.These passages describe the questions asked; they do not report participants’ answers.Source: §3.1 · pp. 13–14

3.2 Future Use of Eye Tracking Data

  • Potential uses included early detection of diseases, hands-free interaction, visual-target identification, improved interfaces, and more understandable content.The survey listed examples such as dementia or Parkinson’s detection, gaze-based selection, and interface improvements.Source: §3.2 · p. 15
  • Other proposed uses included reading and learning support, stress monitoring, interest identification, activity tracking, and shopping-behavior analysis.These are listed survey scenarios, not reported outcomes about willingness to share.Source: §3.2 · p. 15

3.3 Sharing Eye Tracking Data

  • The survey asked whether trust and willingness to share varied by operator or owner, including government agencies, companies, home cloud, and company intranet.It also asked whether participants would share data in exchange for benefits such as shopping assistance or activity logging.Source: §3.3 · p. 16
  • Sharing questions varied collection conditions, including public, private, or constrained settings, VR/AR, indoor or outdoor locations, and restricted recording duration.The survey also asked about sharing for benefits such as shopping assistance or activity logging.Source: §3.3 · pp. 16–18
  • Sharing questions varied collection purpose and duration, as well as emotion, application category, activity, and device context.The prompts included application-specific purposes, unspecified versus half-year durations, and multiple contexts for collection.Source: §3.3 · pp. 19–21
  • The survey also asked about sharing different combinations of gaze or pupil behavior, scene video, and eye video content.Respondents were asked about sharing each content type alone and in combinations.Source: §3.3 · p. 22
  • An anonymous-identity prompt asked whether respondents would hide additional personal information, including location, health, profile image, age, and city.The examples also included occupation and employment.Source: §3.3 · p. 22

3.4 Eye Tracking Data Representations

  • Representation questions covered raw gaze or pupil positions, fixation and saccade statistics, scan paths, gaze plots, heatmaps, and aggregated features.The prompts also included fixation-point and area-of-interest statistics.Source: §3.4 · p. 23
  • A separate hypothetical asked whether respondents would share these representations if the data were modified to be anonymous and indistinguishable from another user’s.The question was posed for raw gaze or pupil positions and the other listed representations.Source: §3.4 · p. 24
  • The survey also identifies behavioral measures such as average blinking rate, fixation duration, and ratio of saccadic movements.Source: §3.4 · p. 24
  • The scene-video question lists whole-recording content, fixation frames, gaze stripes, and tiny gaze-centered image patches as sharing options.The patch options are framed for either a whole recording or each fixation.Source: §3.4 · p. 25
  • Another scene-video option shows the whole scene with surrounding areas blurred and the object of interest clear.Source: §3.4 · p. 25
  • The eye-camera question contrasts video showing the iris and facial expressions, blurred facial expressions, or only the pupil center.Source: §3.4 · p. 25

4 EYE MOVEMENT FEATURE EXTRACTION

The feature extraction section describes 52 eye-movement features across several signal types and scan paths, extracted using overlapping 30-second windows. It also encodes saccades as characters in words of length n.

4 EYE MOVEMENT FEATURE EXTRACTION

  • Each saccade is encoded as a character that forms words of length n, following Bulling et al. 2011.Source: §4 · p. 26
  • Features were extracted in 30-second sliding windows with a 0.5-second step size.Source: §4 · p. 26
  • The extraction comprises 52 eye-movement features spanning fixations, saccades, blinks, pupil diameter, and scan paths.Table 1 gives the number of features per category.Source: Table 1 · p. 26, §4 · p. 26
Loading 1812.08000v3…