Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

Privacy Enhancement for Gaze Data Using a Noise-Infused Autoencoder

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A latent-noise autoencoder can strip gaze data of its biometric identity while preserving its usefulness for benign prediction tasks.

desk verdict A plausible gaze-privacy mechanism that I cannot evaluate because the supplied full text is unreadable; the abstract overclaims 'prevents' where the evidence can only support 'reduces.' read the letter →

arxiv 2508.10918 v1 pith:HMA6ZXSB submitted 2025-08-01 cs.CV cs.HC

Samantha Aziz , Oleg Komogortsev This is my paper · ORCID
classification cs.CVcs.HC
keywords gazedataprivacybiometricre-identificationlatent-noiseautoencoderprivacy-utilitytrade-offpredictioneyetrackingphysiologicalplausibility
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that gaze data can be made safe to share by routing it through an autoencoder that adds noise to the compressed latent code before reconstructing the signal. The perturbed gaze traces significantly reduce cross-session biometric re-identification, the paper reports, while gaze prediction tasks lose only a small amount of accuracy. The authors frame the result as a privacy-utility trade-off and argue that the key advantage over earlier methods is that the output remains physiologically plausible gaze behavior. If the claim holds, eye-tracking systems in gaming, virtual reality, and assistive technology can store or release gaze data without exposing the identity of the user.

What carries the argument

The central object is a noise-infused autoencoder, a neural network that maps gaze traces into a low-dimensional latent code, perturbs that code with random noise, and reconstructs a gaze signal from the perturbed code. The latent noise is the privacy mechanism: it disrupts the identity-specific regularities that biometric identification exploits while leaving intact the broader gaze statistics that benign prediction tasks depend on. The evaluation sets the two sides of the trade-off as biometric identification accuracy and gaze prediction accuracy, with the physiological plausibility of outputs as the factor that keeps the two sides in balance.

What would settle it

Train a re-identification model on the privacy-transformed gaze traces themselves and run it across sessions with a large gallery and a realistic time gap; if identification accuracy returns to a level close to the original, the claimed privacy protection does not survive a determined attacker.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the latent space of an autoencoder is the right place to inject privacy noise for gaze signals. An encoder compresses a gaze trace into a compact code, random noise is added to that code, and a decoder produces a modified trace that no longer carries the identity-specific structure that biometric identification relies on. The paper reports a significant drop in biometric identifiability across play sessions together with minimal degradation on gaze prediction, and it attributes the favorable trade-off to the physiological plausibility of the reconstructed traces. The intended consequence is a practical privacy mechanism: gaze data can be transformed once and then used for benign purposes without enabling re-identification.

Load-bearing premise

The load-bearing premise is that the evaluation protocol measures the real threat: the biometric identification task must reflect the re-identification risk an attacker actually faces across play sessions, and the gaze prediction task must reflect the benign uses that actually matter.

Editorial extensions

If this is right

  • Gaze-data holders could release transformed traces that resist cross-session re-identification while still supporting benign gaze-prediction tasks.
  • The latent noise level becomes a tunable dial, letting a system choose where it sits on the privacy-utility frontier.
  • Because outputs remain physiologically plausible, benign downstream models are less likely to be disrupted by reconstruction artifacts than by input-level perturbations.
  • The approach can be applied at the point of collection, so raw identifiable gaze never leaves the device.
  • Privacy becomes a transformation problem rather than a suppression problem: the data can stay useful without exposing the person.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's threat model assumes an attacker who runs the same kind of identification benchmark; a stronger adversary who trains an identifier on the transformed traces could partially undo the privacy gain, and that scenario is not settled by the reported numbers.
  • The same latent-noise design should transfer to other continuous biometric signals, such as ECG, EEG, or gait, wherever an autoencoder can separate identity-specific from task-relevant variation.
  • A natural follow-up is to calibrate the latent noise to a formal privacy definition such as differential privacy, which would turn the empirical trade-off into a provable bound.
  • A quantitative metric for physiological plausibility would let future work test directly whether realism is what preserves utility, rather than treating it as a stated property.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes a privacy-enhancing mechanism for gaze signals based on a latent-noise autoencoder. The abstract claims that injecting noise into the latent space of an autoencoder prevents users from being re-identified across play sessions without their consent, while preserving the utility of the data for benign tasks. The authors report evaluating privacy-utility trade-offs across biometric identification and gaze prediction tasks, asserting a significant reduction in biometric identifiability with minimal utility degradation, and further claim that the framework retains physiologically plausible gaze patterns. Because the supplied full text is unreadable mojibake, no equations, datasets, baselines, results, or ablations are accessible in the submission, leaving the central empirical claim entirely unverifiable from the available material.

Significance. If the claimed results are correct, the paper would make a valuable contribution to privacy in gaze-based systems by demonstrating that a latent-noise autoencoder can serve as a practical, tunable privacy mechanism with a favorable privacy-utility trade-off. The problem is well motivated and the method is plausible. However, the significance cannot currently be assessed from the submission: the full text is corrupted, and the abstract provides no quantitative evidence, no dataset names, no baseline comparisons, and no description of the evaluation protocol. The paper does not, in the available material, ship reproducible code or machine-checked proofs, so the central claim rests entirely on the unreadable full text.

major comments (4)
  1. [Full text (as supplied)] The full text supplied to the referee is corrupted mojibake; no equations, tables, figures, dataset names, or numerical results are legible. The central claim that the proposed autoencoder significantly reduces biometric identifiability with minimal utility degradation is therefore unverifiable from this submission. The authors must resubmit a readable version with the complete methodology and all empirical results before the manuscript can be reviewed.
  2. [Abstract, first sentence] The abstract states that the mechanism "prevents users from being re-identified," but the evaluation described in the abstract can at most support a reduction in identifiability. This wording overstates the likely result and should be corrected, since the paper has not demonstrated a complete absence of re-identification risk.
  3. [Abstract, evaluation description] The abstract reports no datasets, baseline methods, identification accuracy, false-acceptance rates, utility metrics, or noise-level settings. The claimed privacy-utility trade-off is the central result, and without specifying the biometric identification protocol (identification model, gallery size, session gap, operating point) and the gaze prediction task, the claim is not reproducible and its real-world validity cannot be assessed. A concrete test is to report the privacy-utility curve over noise levels under a cross-session identification protocol.
  4. [Abstract, physiologically plausible gaze patterns] The claim that the framework retains physiologically plausible gaze patterns is not supported by any visible evaluation criterion. The authors should specify the plausibility measure (e.g., comparison with known saccade and fixation statistics, perceptual evaluation, or a quantitative plausibility metric) and report the corresponding results, since this claim is presented as an advantage over prior methods.
minor comments (3)
  1. [Abstract] The phrase "usable and effective mechanism" is vague; consider specifying the exact privacy guarantee provided (e.g., reduction to chance-level identification or a concrete upper bound on re-identification accuracy).
  2. [Abstract, 'without their consent'] The phrase "without their consent" is ambiguous: is the privacy mechanism applied by the data processor, by the user's client, or by the platform? Clarifying the trust model would help position the contribution.
  3. [Footer (arXiv:2508.10919)] The extracted footer shows an arXiv identifier different from the manuscript ID (2508.10918); please check whether this is an artifact of extraction or a genuine mismatch in the submission metadata.
Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

This ledger is derived from the abstract alone because the submitted full text is unreadable, being corrupted into mojibake. The free parameters listed are those implied by the method description; the axioms are the domain assumptions that must hold for the claimed privacy-utility trade-off to be real. No new physical or conceptual entities are introduced.

free parameters (2)
  • latent noise magnitude
    The abstract describes a noise-infused autoencoder; the amount of noise injected in the latent space is the dial that sets the privacy-utility trade-off, but no value or selection rule is reported in the abstract.
  • autoencoder architecture and training hyperparameters
    Bottleneck width, loss weighting, and training schedule are not reported in the abstract, and they would be required to reproduce the claimed trade-off.
assumptions (3)
  • domain assumption Identity-discriminative information in gaze is separable from task-relevant information, so latent noise can remove the former while preserving the latter.
    This is the core modeling premise of the method: that the autoencoder can decorrelate identity from gaze content rather than destroying both. The abstract asserts the outcome, not the premise.
  • domain assumption Biometric identification accuracy on the evaluation protocol is a valid estimate of cross-session re-identification risk.
    The abstract equates reduced biometric identifiability with protection against re-identification across play sessions; the fidelity of that proxy, in terms of attacker model, gallery size, and session gaps, is assumed and never specified in the abstract.
  • standard math Standard autoencoder training with stochastic noise converges to a useful reconstruction of gaze signals.
    Background assumption of any trained autoencoder; standard machine learning practice and not the paper's contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Privacy Enhancement for Gaze Data Using a Noise-Infused Autoencoder." pith.science (2026). https://pith.science/paper/HMA6ZXSB

@misc{pith2026250810918,
  author       = {Pith},
  title        = {Pith review of: Privacy Enhancement for Gaze Data Using a Noise-Infused Autoencoder},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HMA6ZXSB}},
  note         = {Machine review of arXiv:2508.10918}
}
read the original abstract

We present a privacy-enhancing mechanism for gaze signals using a latent-noise autoencoder that prevents users from being re-identified across play sessions without their consent, while retaining the usability of the data for benign tasks. We evaluate privacy-utility trade-offs across biometric identification and gaze prediction tasks, showing that our approach significantly reduces biometric identifiability with minimal utility degradation. Unlike prior methods in this direction, our framework retains physiologically plausible gaze patterns suitable for downstream use, which produces favorable privacy-utility trade-off. This work advances privacy in gaze-based systems by providing a usable and effective mechanism for protecting sensitive gaze data.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Privatization of Synthetic Gaze: Attenuating State Signatures in Diffusion-Generated Eye Movements

    cs.HC 2026-01 conditional novelty 4.0 of 10

    Diffusion-generated synthetic gaze shows weaker, less stable correlations with self-reported fatigue and difficulty than real gaze, suggesting partial state-level privatization.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Practical perception-based evaluation of gaze prediction for gaze contingent rendering

    Samantha Aziz, Dillon J Lohr, Razvan Stefanescu, and Oleg Komogortsev. Practical perception-based evaluation of gaze prediction for gaze contingent rendering. Proceedings of the ACM on Human-Computer Interaction , 7(ETRA):1--17, 2023

    work page 2023

  2. [2]

    Baji\' c

    Bardia Azizian and Ivan V. Baji\' c . Privacy-preserving autoencoder for collaborative object detection. Trans. Img. Proc. , 33:4937–4951, January 2024

    work page 2024

  3. [3]

    Joint attention simulation using eye-tracking and virtual humans

    Matthieu Courgeon, Gilles Rautureau, Jean-Claude Martin, and Ouriel Grynszpan. Joint attention simulation using eye-tracking and virtual humans. IEEE Transactions on Affective Computing , 5(3):238--250, 2014

    work page 2014

  4. [4]

    Soft- DTW : a differentiable loss function for time-series

    Marco Cuturi and Mathieu Blondel. Soft- DTW : a differentiable loss function for time-series. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning , volume 70 of Proceedings of Machine Learning Research , pages 894--903. PMLR, 06--11 Aug 2017

    work page 2017

  5. [5]

    A privacy-preserving approach to streaming eye-tracking data

    Brendan David-John, Diane Hosfelt, Kevin Butler, and Eakta Jain. A privacy-preserving approach to streaming eye-tracking data. IEEE Transactions on Visualization and Computer Graphics , PP:1--1, 03 2021

    work page 2021

  6. [6]

    Differentially private recommender framework with dual semi-autoencoder

    Yang Deng, Wang Zhou, Amin Ul Haq, Sultan Ahmad, and Alia Tabassum. Differentially private recommender framework with dual semi-autoencoder. Expert Syst. Appl. , 260(C), February 2025

    work page 2025

  7. [7]

    Reinforcement learning for the privacy preservation and manipulation of eye tracking data

    Wolfgang Fuhl, Efe Bozkir, and Enkelejda Kasneci. Reinforcement learning for the privacy preservation and manipulation of eye tracking data. In Igor Farka s , Paolo Masulli, Sebastian Otte, and Stefan Wermter, editors, Artificial Neural Networks and Machine Learning -- ICANN 2021 , pages 595--607, Cham, 2021. Springer International Publishing

    work page 2021

  8. [8]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015

    2015

Show all 30 references
  1. [9]

    Gazebase, a large-scale, multi-stimulus, longitudinal eye movement dataset

    Henry Griffith, Dillon Lohr, Evgeny Abdulin, and Oleg Komogortsev. Gazebase, a large-scale, multi-stimulus, longitudinal eye movement dataset. Scientific Data , 8(1):184, Jul 2021

    2021

  2. [10]

    Otus: A gaze model-based privacy control framework for eye tracking applications

    Miao Hu, Zhenxiao Luo, Yipeng Zhou, Xuezheng Liu, and Di Wu. Otus: A gaze model-based privacy control framework for eye tracking applications. In IEEE INFOCOM 2022 - IEEE Conference on Computer Communications , pages 560--569, 2022

    2022

  3. [11]

    An eye-tracking assistive device improves the quality of life for als patients and reduces the caregivers’ burden

    Chi-Shin Hwang, Ho-Hsiu Weng, Li-Fen Wang, Chon-Haw Tsai, and Hao-Teng Chang. An eye-tracking assistive device improves the quality of life for als patients and reduces the caregivers’ burden. Journal of motor behavior , 46(4):233--238, 2014

    2014

  4. [12]

    Jamshidi, Hadi Veisi, Mohammad M

    Mohammad A. Jamshidi, Hadi Veisi, Mohammad M. Mojahedian, and Mohammad R. Aref. Adjustable privacy using autoencoder-based learning structure. Neurocomputing , 566:127043, 2024

    2024

  5. [13]

    Eye tracking cognitive load using pupil diameter and microsaccades with fixed gaze

    Krzysztof Krejtz, Andrew T Duchowski, Anna Niedzielska, Cezary Biele, and Izabela Krejtz. Eye tracking cognitive load using pupil diameter and microsaccades with fixed gaze. PloS one , 13(9):e0203629, 2018

    2018

  6. [14]

    o ger, Otto Hans-Martin Lutz, and Florian M \

    Jacob Leon Kr \"o ger, Otto Hans-Martin Lutz, and Florian M \"u ller. What does your gaze reveal about you? on the privacy implications of eye tracking. In IFIP International Summer School on Privacy and Identity Management , pages 226--241. Springer, 2020

    2020

  7. [15]

    Differences in eye movement range based on age and gaze direction

    Won June Lee, Ji Hong Kim, Yong Un Shin, Sunjin Hwang, and Han Woong Lim. Differences in eye movement range based on age and gaze direction. Eye , 33(7):1145--1151, 2019

    2019

  8. [16]

    Kaleido : Real-Time privacy control for Eye-Tracking systems

    Jingjie Li, Amrita Roy Chowdhury, Kassem Fawaz, and Younghyun Kim. Kaleido : Real-Time privacy control for Eye-Tracking systems. In 30th USENIX Security Symposium (USENIX Security 21) , pages 1793--1810. USENIX Association, August 2021

  9. [17]

    Differential privacy for eye-tracking data

    Ao Liu, Lirong Xia, Andrew Duchowski, Reynold Bailey, Kenneth Holmqvist, and Eakta Jain. Differential privacy for eye-tracking data. In Proceedings of the 11th ACM Symposium on Eye Tracking Research & Applications , ETRA '19, New York, NY, USA, 2019. Association for Computing ...

    2019

  10. [18]

    Enhanced embedded autoencoders: An attribute-preserving face de-identification framework

    Jianqi Liu, Zhiwei Zhao, Pan Li, Geyong Min, and Huiyong Li. Enhanced embedded autoencoders: An attribute-preserving face de-identification framework. IEEE Internet of Things Journal , 10(11):9438--9452, 2023

    2023

  11. [19]

    Differentially private recommender system with autoencoders

    Xiaoqian Liu, Qianmu Li, Zhen Ni, and Jun Hou. Differentially private recommender system with autoencoders. In 2019 International Conference on Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing (CP...

    2019

  12. [20]

    Komogortsev

    Dillon Lohr and Oleg V. Komogortsev. Eye know you too: Toward viable end-to-end eye movement biometrics for user authentication. IEEE Transactions on Information Forensics and Security , 17:3151--3164, 2022

    2022

  13. [21]

    Establishing a baseline for gaze-driven authentication performance in vr: A breadth-first investigation on a very large dataset

    Dillon Lohr, Michael J Proulx, and Oleg Komogortsev. Establishing a baseline for gaze-driven authentication performance in vr: A breadth-first investigation on a very large dataset. In 2024 IEEE International Joint Conference on Biometrics (IJCB) , pages 1--10. IEEE, 2024

    2024

  14. [22]

    Uncertainty-autoencoder-based privacy and utility preserving data type conscious transformation

    Bishwas Mandal, George Amariucai, and Shuangqing Wei. Uncertainty-autoencoder-based privacy and utility preserving data type conscious transformation. In 2022 International Joint Conference on Neural Networks (IJCNN) , pages 1--8, 2022

    2022

  15. [23]

    Raschka, Anoop M

    Vahid Mirjalili, S. Raschka, Anoop M. Namboodiri, and Arun Ross. Semi-adversarial networks: Convolutional autoencoders for imparting privacy to face images. 2018 International Conference on Biometrics (ICB) , pages 82--89, 2017

    2018

  16. [24]

    Towards foveated rendering for gaze-tracked virtual reality

    Anjul Patney, Marco Salvi, Joohwan Kim, Anton Kaplanyan, Chris Wyman, Nir Benty, David Luebke, and Aaron Lefohn. Towards foveated rendering for gaze-tracked virtual reality. ACM Trans. Graph. , 35(6), dec 2016

    2016

  17. [25]

    Perero-Codosero, Fernando M

    Juan M. Perero-Codosero, Fernando M. Espinoza-Cuadros, and Luis A. Hernández-Gómez. X-vector anonymization using autoencoders and adversarial training for preserving speech privacy. Computer Speech & Language , 74:101351, 2022

    2022

  18. [26]

    The eye in extended reality: A survey on gaze interaction and eye tracking in head-worn extended reality

    Alexander Plopski, Teresa Hirzle, Nahal Norouzi, Long Qian, Gerd Bruder, and Tobias Langlotz. The eye in extended reality: A survey on gaze interaction and eye tracking in head-worn extended reality. ACM Computing Surveys (CSUR) , 55(3):1--39, 2022

    2022

  19. [27]

    J\" a ger

    Paul Prasse, David Robert Reich, Silvia Makowski, Seoyoung Ahn, Tobias Scheffer, and Lena A. J\" a ger. Sp-eyegan: Generating synthetic eye movement data with generative adversarial networks. In Proceedings of the 2023 Symposium on Eye Tracking Research and Applications , ETRA...

    2023

  20. [28]

    Recommender systems based on autoencoder and differential privacy

    Jiahui Ren, Xian Xu, Zhihuan Yao, and Huiqun Yu. Recommender systems based on autoencoder and differential privacy. In 2019 IEEE 43rd Annual Computer Software and Applications Conference (COMPSAC) , volume 1, pages 358--363, 2019

    2019

  21. [29]

    Privacy-aware eye tracking using differential privacy

    Julian Steil, Inken Hagestedt, Michael Xuelin Huang, and Andreas Bulling. Privacy-aware eye tracking using differential privacy. In Proceedings of the 11th ACM Symposium on Eye Tracking Research & Applications , ETRA '19, New York, NY, USA, 2019. Association for Computing Machinery

    2019

  22. [30]

    Proulx, Sai Deep Tetali, Kevin Butler, and Eakta Jain

    Ethan Wilson, Azim Ibragimov, Michael J. Proulx, Sai Deep Tetali, Kevin Butler, and Eakta Jain. Privacy-preserving gaze data streaming in immersive interactive virtual reality: Robustness and user experience, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.