pith. sign in

arxiv: 1906.11979 · v1 · pith:2C4DJ3WPnew · submitted 2019-06-27 · 💻 cs.CV · cs.CR· cs.LG· eess.IV

A Utility-Preserving GAN for Face Obscuration

Hanxiang Hao , David G\"uera , Amy R. Reibman , Edward J. Delp This is my paper

Pith reviewed 2026-05-25 14:31 UTC · model grok-4.3

classification 💻 cs.CV cs.CRcs.LGeess.IV
keywords face obscurationgenerative adversarial networkprivacy protectionutility preservationface anonymizationidentity concealmentdeep learningimage synthesis
0
0 comments X

The pith

A generative model called UP-GAN obscures facial identity while preserving age, gender, skin tone, pose, and expression.

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

The paper introduces UP-GAN, a generative adversarial network that produces face images with identity hidden but with non-identifying attributes left intact. Traditional obscuration techniques such as blurring and pixelation no longer guarantee privacy because deep learning models can often recover the original identity. A reader would care if the separation of identity from utility attributes holds, because it would let public images retain value for tasks that depend on age, gender, pose, or expression without exposing who the person is.

Core claim

The authors present UP-GAN as a utility-preserving generative model that conceals facial identity through adversarial training while retaining utility attributes, and they report that it outperforms prior obscuration methods on both identity concealment and utility retention.

What carries the argument

UP-GAN, a generative adversarial network that separates identity features from utility attributes during image synthesis.

If this is right

  • Obscured images from sources such as news or mapping services can retain utility for analysis while reducing re-identification risk.
  • Utility attributes including age, gender, skin tone, pose, and expression remain measurable after obscuration.
  • The method provides stronger protection than Gaussian blurring or pixelation against modern re-identification attacks.
  • The same separation principle could apply to other image datasets that require both privacy and downstream utility.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Real-time versions of this model could be inserted into video pipelines to anonymize faces on the fly.
  • Training recognition systems on UP-GAN outputs might reduce privacy leakage in public datasets.
  • The approach raises the question of whether similar separation can be achieved for non-facial identifiers such as gait or clothing.

Load-bearing premise

Facial identity can be separated from utility attributes such as age and gender so that a generative model can hide one without damaging the others.

What would settle it

An identity classifier that achieves high accuracy on UP-GAN obscured faces at rates similar to its accuracy on the original unprocessed faces.

Figures

Figures reproduced from arXiv: 1906.11979 by Amy R. Reibman, David G\"uera, Edward J. Delp, Hanxiang Hao.

Figure 1
Figure 1. Figure 1: Obscuration effect of the proposed method. First row: original faces; second row: obscured faces. facial information, while preserving the features that do not convey identity. The proposed method, utility-preserving GAN (UP-GAN), aims to provide an effective obscuration by generating faces that only depend on the non-identifiable facial features. In this work, we define utility as the facial properties su… view at source ↗
Figure 3
Figure 3. Figure 3: Example of the augmented face with elastic distortion and random rotation (left) and its binary mask (right). compared to Gaussian blurring, pixelation, k-same method and k-same-net method. 4.1. Datasets The UTKFace dataset (Zhang et al., 2017) contains 23,708 images with annotations of 68-point facial landmarks and attributes of age, gender, and skin tone. To obscure the identifiable information present i… view at source ↗
Figure 4
Figure 4. Figure 4: Generated faces with different loss functions. ceptual network and binary mask. We can also see that, compared to the original face, the generated face with ad￾versarial loss and L2 reconstruction loss can preserve the facial utility. However, the facial details such as the outlines are partially missing. By adding the mask loss, we can en￾hance the facial boundary, like the cheek and chin. If we add the p… view at source ↗
Figure 5
Figure 5. Figure 5: Examples of obscured faces. Top row: original image, k-same (k = 10) and Gaussian blurring (kernel sizes: 5, 15 and 25). Bottom row: k-same-net, UP-GAN and pixelation (pixel sizes: 5, 15 and 25). sian blurring with kernel size 5 and 15 fail to provide an effective obscuration, while all other methods achieve good performance. For the threat model II, the obscuration perfor￾mance degrades for all methods, w… view at source ↗
read the original abstract

From TV news to Google StreetView, face obscuration has been used for privacy protection. Due to recent advances in the field of deep learning, obscuration methods such as Gaussian blurring and pixelation are not guaranteed to conceal identity. In this paper, we propose a utility-preserving generative model, UP-GAN, that is able to provide an effective face obscuration, while preserving facial utility. By utility-preserving we mean preserving facial features that do not reveal identity, such as age, gender, skin tone, pose, and expression. We show that the proposed method achieves the best performance in terms of obscuration and utility preservation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes UP-GAN, a utility-preserving generative adversarial network for face obscuration. The method is intended to conceal identity while retaining non-identifying attributes (age, gender, skin tone, pose, expression). The central claim is that UP-GAN achieves the best performance among compared methods on both obscuration effectiveness and utility preservation.

Significance. If the empirical claims are substantiated with robust evaluation, the result would be relevant to privacy-preserving computer vision pipelines (e.g., Street View, broadcast media) where simple blurring or pixelation is now known to be insufficient against modern recognizers. The approach directly targets the privacy-utility trade-off via an adversarial formulation rather than post-hoc filtering.

major comments (1)
  1. [Experimental results / Evaluation protocol] The obscuration claim rests on lowered accuracy of one or more face-recognition networks on the generated images. Because identity and utility attributes (pose, expression) are statistically entangled, any generator that preserves the latter can still leak identity to a recognizer whose feature space was not explicitly penalized. The manuscript reports results only against the recognizer(s) used during training or architecturally similar models; no results are given for independent, stronger, or disjoint recognizers (different backbone or training corpus). This leaves the central claim vulnerable to the concern that measured success does not transfer to an external adversary.
minor comments (2)
  1. [Abstract] The abstract asserts 'best performance' without any quantitative metrics, baselines, or dataset names; while the full experimental section presumably supplies these, the summary paragraph should at minimum indicate the evaluation protocol and primary numbers.
  2. [Method] Notation for the utility and identity losses is introduced without an explicit equation reference or table summarizing all loss terms and their weighting coefficients.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the single major comment below.

read point-by-point responses

  1. Referee: [Experimental results / Evaluation protocol] The obscuration claim rests on lowered accuracy of one or more face-recognition networks on the generated images. Because identity and utility attributes (pose, expression) are statistically entangled, any generator that preserves the latter can still leak identity to a recognizer whose feature space was not explicitly penalized. The manuscript reports results only against the recognizer(s) used during training or architecturally similar models; no results are given for independent, stronger, or disjoint recognizers (different backbone or training corpus). This leaves the central claim vulnerable to the concern that measured success does not transfer to an external adversary.

    Authors: We agree that the evaluation would be strengthened by results on independent recognizers with different backbones or training corpora. The adversarial objective in UP-GAN is formulated to penalize identity leakage while preserving utility attributes, and the reported experiments already include multiple recognizer architectures. Nevertheless, the referee's concern about generalization is valid. In the revised manuscript we will add quantitative results against at least two additional, disjoint face-recognition models (different backbone and training set) to demonstrate that the measured obscuration transfers beyond the training recognizer. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical evaluation only

full rationale

The paper proposes UP-GAN as an empirical generative model and reports performance on obscuration and utility metrics. No derivation chain, equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described structure. The central claim rests on experimental comparisons rather than any self-definitional reduction or ansatz smuggled via prior work. This is the expected outcome for a standard applied ML paper without theoretical derivations.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient detail to identify specific free parameters, axioms, or invented entities beyond the high-level model proposal.

pith-pipeline@v0.9.0 · 5645 in / 923 out tokens · 35635 ms · 2026-05-25T14:31:03.188332+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 6 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

    work page

  2. [2]

    Face aging with conditional generative adversarial networks

    Antipov, G., Baccouche, M., and Dugelay, J.-L. Face aging with conditional generative adversarial networks. Proceedings of the IEEE International Conference on Image Processing, pp.\ 2089--2093, Sep. 2017. URL https://doi.org/10.1109/ICIP.2017.8296650. Beijing, China

    work page doi:10.1109/icip.2017.8296650 2089

  3. [3]

    Bitouk, D., Kumar, N., Dhillon, S., Belhumeur, P., and Nayar, S. K. Face swapping: Automatically replacing faces in photographs. ACM Transactions on Graphics, 27 0 (3): 0 39:1--39:8, 2008. URL https://doi.org/10.1145/1360612.1360638

    work page doi:10.1145/1360612.1360638 2008

  4. [4]

    F., Edwards , G

    Cootes , T. F., Edwards , G. J., and Taylor , C. J. Active appearance models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23 0 (6): 0 484--498, June 2001. URL https://doi.org/10.1109/34.927467

    work page doi:10.1109/34.927467 2001

  5. [5]

    T., Tatarchenko, M., and Brox, T

    Dosovitskiy, A., Springenberg, J. T., Tatarchenko, M., and Brox, T. Learning to generate chairs, tables and cars with convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39 0 (4): 0 692--705, April 2017. URL https://doi.org/10.1109/TPAMI.2016.2567384

    work page doi:10.1109/tpami.2016.2567384 2017

  6. [6]

    Garp-face: Balancing privacy protection and utility preservation in face de-identification

    Du, L., Yi, M., Blasch, E., and Ling, H. Garp-face: Balancing privacy protection and utility preservation in face de-identification. Proceedings of the IEEE International Joint Conference on Biometrics, pp.\ 1--8, September 2014. URL https://doi.org/10.1109/BTAS.2014.6996249. Clearwater, FL

    work page doi:10.1109/btas.2014.6996249 2014

  7. [7]

    Generative Adversarial Networks

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial nets. Advances in Neural Information Processing Systems, pp.\ 2672--2680, December 2014. URL https://arxiv.org/abs/1406.2661. Montr\'eal, Canada

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  8. [8]

    Integrating utility into face de-identification

    Gross, R., Airoldo, E., Malin, B., and Sweeney, L. Integrating utility into face de-identification. Proceedings of the International Workshop on Privacy Enhancing Technologies, pp.\ 227--242, May 2005. URL https://doi.org/10.1007/11767831_15. Cavtat, Croatia

    work page doi:10.1007/11767831_15 2005

  9. [9]

    and Delp , E

    G\" u era, D. and Delp, E. J. Deepfake video detection using recurrent neural networks. Proceedings of the IEEE International Conference on Advanced Video and Signal Based Surveillance, pp.\ 1--6, Nov. 2018. URL https://doi.org/10.1109/AVSS.2018.8639163. Auckland, New Zealand

    work page doi:10.1109/avss.2018.8639163 2018

  10. [10]

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, pp.\ 6629--6640, 2017. URL http://arxiv.org/abs/1706.08500. Long Beach, CA

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    Perceptual losses for real-time style transfer and super-resolution

    Johnson, J., Alahi, A., and Fei-Fei, L. Perceptual losses for real-time style transfer and super-resolution. European Conference on Computer Vision, pp.\ 694--711, 2016. URL https://doi.org/10.1007/978-3-319-46475-6_43. Amsterdam, Netherlands

    work page doi:10.1007/978-3-319-46475-6_43 2016

  12. [12]

    King, D. E. Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research, 10: 0 1755--1758, December 2009. URL http://dl.acm.org/citation.cfm?id=1577069.1755843

    work page arXiv 2009

  13. [13]

    Fast face-swap using convolutional neural networks

    Korshunova, I., Shi, W., Dambre, J., and Theis, L. Fast face-swap using convolutional neural networks. Proceedings of the IEEE International Conference on Computer Vision, pp.\ 3697--3705, Oct. 2017. URL https://doi.org/10.1109/ICCV.2017.397. Venice, Italy

    work page doi:10.1109/iccv.2017.397 2017

  14. [14]

    Attribute-Guided Face Generation Using Conditional CycleGAN

    Lu, Y., Tai, Y.-W., and Tang, C.-K. Attribute-guided face generation using conditional cyclegan. arXiv:1705.09966v2, Nov. 2018. URL https://arxiv.org/abs/1705.09966

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    Defeating Image Obfuscation with Deep Learning

    McPherson, R., Shokri, R., and Shmatikov, V. Defeating image obfuscation with deep learning. arXiv:1609.00408v2, September 2016. URL https://arxiv.org/abs/1609.00408

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  16. [16]

    Emer s i c , S truc, V., and Peer, P

    Meden, B., Z . Emer s i c , S truc, V., and Peer, P. k-same-net: k-anonymity with generative deep neural networks for face deidentification. Entropy, 20 0 (1), January 2018. URL https://doi.org/10.3390/e20010060

    work page doi:10.3390/e20010060 2018

  17. [17]

    Conditional Generative Adversarial Nets

    Mirza, M. and Osindero, S. Conditional generative adversarial nets. arXiv:1411.1784v1, November 2014. URL https://arxiv.org/abs/1411.1784

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  18. [18]

    M., Sweeney , L., and Malin , B

    Newton , E. M., Sweeney , L., and Malin , B. Preserving privacy by de-identifying face images. IEEE Transactions on Knowledge and Data Engineering, pp.\ 232--243, Feburary 2005. URL https://doi.org/10.1109/TKDE.2005.32

    work page doi:10.1109/tkde.2005.32 2005

  19. [19]

    and Winkler, S

    Ng, H. and Winkler, S. A data-driven approach to cleaning large face datasets. Proceedings of the IEEE International Conference on Image Processing, pp.\ 343--347, October 2014. URL https://doi.org/10.1109/ICIP.2014.7025068. Paris, France

    work page doi:10.1109/icip.2014.7025068 2014

  20. [20]

    Poisson image editing

    P \'e rez, P., Gangnet, M., and Blake, A. Poisson image editing. ACM Transactions on Graphics, 22 0 (3): 0 313--318, July 2003. URL http://doi.acm.org/10.1145/882262.882269

    work page doi:10.1145/882262.882269 2003

  21. [21]

    ImageNet Large Scale Visual Recognition Challenge,

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. ImageNet Large Scale Visual Recognition Challenge . International Journal of Computer Vision, 115 0 (3): 0 211--252, 2015. URL https://doi.org/10.1007/s11263-015-0816-y

    work page doi:10.1007/s11263-015-0816-y 2015

  22. [22]

    and Sweeney, L

    Samarati, P. and Sweeney, L. Protecting privacy when disclosing information: k-anonymity and its enforcement through generalization and suppression. Technical Report, 1998. Harvard Data Privacy Laboratory

    work page 1998

  23. [23]

    Y., Steinkraus, D., and Platt, J

    Simard, P. Y., Steinkraus, D., and Platt, J. C. Best practices for convolutional neural networks applied to visual document analysis. Proceedings of the IEEE International Conference on Document Analysis and Recognition, 2: 0 958--963, August 2003. URL https://doi.org/10.1109/ICDAR.2003.1227801. Edinburgh, Scotland

    work page doi:10.1109/icdar.2003.1227801 2003

  24. [24]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. Proceedings of the International Conference on Learning Representations, May 2015. URL http://arxiv.org/abs/1409.1556. San Diego, CA

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  25. [25]

    V., Schiele , B., and Fritz , M

    Sun , Q., Ma , L., Joon Oh , S., Gool , L. V., Schiele , B., and Fritz , M. Natural and effective obfuscation by head inpainting. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 5050--5059, June 2018. URL https://doi.org/10.1109/CVPR.2018.00530. Salt Lake City, UT

    work page doi:10.1109/cvpr.2018.00530 2018

  26. [26]

    Privacy-protective-gan for privacy preserving face de-identification

    Wu, Y., Yang, F., Xu, Y., and Ling, H. Privacy-protective-gan for privacy preserving face de-identification. Journal of Computer Science and Technology, pp.\ 47--60, January 2019. URL https://doi.org/10.1007/s11390-019-1898-8. Beijing, China

    work page doi:10.1007/s11390-019-1898-8 2019

  27. [27]

    Age progression/regression by conditional adversarial autoencoder

    Zhang, Z., Song, Y., and Qi, H. Age progression/regression by conditional adversarial autoencoder. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 4352--4360, 2017. URL https://doi.org/10.1109/CVPR.2017.463. Hawaii, HI

    work page doi:10.1109/cvpr.2017.463 2017