Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Empowering Manufacturers with Privacy-Preserving AI Tools: A Case Study in Privacy-Preserving Machine Learning to Solve Real-World Problems

T0 review · 3 major / 4 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 The paper claims a privacy-preserving data-sharing platform lets a manufacturer share proprietary microscope images with researchers, who returned a crystal-counting tool that is 10 times faster than manual counting and accurate enough…

desk verdict A real deployed crystal-counting tool, but the paper's central privacy guarantee is contradicted by its own deployment description. read the letter →

arxiv 2507.01808 v1 pith:NHHPL44K submitted 2025-07-02 cs.CR cs.AIcs.CVcs.ET

Xiaoyu Ji , Jessica Shorland , Joshua Shank , Pascal Delpe-Brice , Latanya Sweeney , Jan Allebach , Ali Shakouri This is my paper
classification cs.CRcs.AIcs.CVcs.ET
keywords privacy-preservingmachinelearningmanufacturingcrystalcountingimageanalysisdatacommonsStarDistqualitycontrolconfidentiality
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

Small and medium manufacturers have factory-floor data that could train useful AI tools, but they refuse to share it because competitors might see it. The paper claims that a privacy-preserving data-sharing platform, MyDataCan, solves this: the manufacturer uploads encrypted images, researchers build and refine a tool without ever seeing raw data, and the finished tool runs back on the platform for daily use. The demonstration case is a crystal-counting tool for food manufacturing, which the paper says was adopted by the manufacturer and cuts per-image analysis time from about five minutes to about thirty seconds, a tenfold speedup, while counting the whole image instead of one ninth of it. If this sharing loop works as described, it would give small and medium manufacturers a practical route to custom AI tools without surrendering proprietary information.

What carries the argument

The argument rests on two coupled mechanisms. The first is MyDataCan, a data-sharing platform that stores each user's files in a private 'can' and, according to the paper, encrypts data throughout upload, transit, analysis, and storage while using homomorphic encryption, federated learning, secure multi-party computation, and zero-knowledge proofs so that neither the platform nor other users can read the data. The second is the crystal counter itself: an adaptive-thresholding and bubble-removal image-processing pipeline for the old camera, and a StarDist model, which is star-convex polygon detection on a U-Net backbone, fine-tuned on new-camera images after dimming the brightest 20 percent of pixels, for the new camera. The platform supplies the trust that makes proprietary data available; the algorithm supplies the value that makes the manufacturer want to use the result.

What would settle it

Watch the actual data flow on the factory floor: if the image bytes leave the platform's encrypted storage and are processed on the cloud compute instance described in the paper in a form the platform operator or cloud provider can read, then the claim that the platform itself cannot access the data is false. A second check is to compare the tool's counts with manual counts on a labeled set of images: if it is not systematically at least as accurate, the accuracy claim fails.

Watch

Extended reading notes

Core claim

On its own terms, this paper's central discovery is that the AI Data Community model works end to end in a real manufacturing setting. A manufacturer that would not share proprietary microscope images was willing to put them in MyDataCan; researchers then developed a crystal counter with two tracks, namely classical image processing for older low-resolution images and a StarDist-based deep-learning model for newer high-resolution images; the tool was packaged as a web app; and the manufacturer adopted it for daily operations. The reported result is not just a faster pipeline: the tool replaces a subjective manual count with a consistent automated count of the full image, and the platform gives the manufacturer confidentiality guarantees throughout the process.

Load-bearing premise

The platform's guarantee that it can never read or decrypt a manufacturer's data is asserted by citing the authors' own prior work, with no implementation details, threat model, or independent security verification offered here.

Editorial extensions

If this is right

  • Other small and medium manufacturers can repeat the same loop, sharing encrypted data, receiving a custom analysis tool, and running it internally without revealing proprietary processes to competitors.
  • The crystal counter can be extended to later manufacturing stages where crystal agglomeration is heavy and no quantitative analysis is currently performed.
  • The image-analysis approach transfers to pharmaceuticals and other particle-processing industries, where crystal size distribution is a quality metric.
  • Because the analyzed image shows each counted crystal outlined, operators can verify the model's output by manual counting, supporting daily trust in the tool.

Reading between the lines

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

  • The paper does not address the fact that its own architecture sends image data through a separately hosted cloud compute instance; if that instance can read the images, the end-to-end confidentiality claim is narrower than stated.
  • A natural test the paper leaves undone is to measure operator-to-operator variance in manual counts before and after deployment, since it claims consistency but reports no variance numbers.
  • The platform model would be more convincing as a general template if it came with an independent security audit or a formal threat model; that is a next step the paper leaves open.
  • A stricter version of the same idea would let competing manufacturers share only derived statistics or model updates, not raw images; the paper demonstrates the weaker, raw-image version.
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

3 major / 4 minor

Summary. This paper presents a case study in which a manufacturer shares proprietary microscope images through the MyDataCan privacy-preserving platform, researchers develop a crystal-counting tool (classical image processing for old camera images and a StarDist-based deep learning model for new camera images), and the tool is deployed as a web application on the factory floor. The paper claims two main contributions: demonstrating a privacy-preserving AI Data Community model for small- and medium-sized manufacturers, and providing a crystal counter that is both faster and more accurate than manual counting. The reported time savings are based on a single comparison, and the accuracy claim is not quantified against ground truth. The privacy guarantees are asserted in Section 3.3.1 and attributed to a prior publication on MyDataCan, but the deployment description in Section 3.3.2 does not describe any cryptographic or enclave mechanisms that would support end-to-end confidentiality.

Significance. If the privacy and accuracy claims were substantiated, this would be a valuable case study showing how small and medium-sized manufacturers can benefit from collaborative machine learning development without exposing proprietary data. The real-world deployment and the documented operator-driven user interface revisions are genuine strengths, as is the detailed description of the image processing pipeline. However, the manuscript does not provide an implementation-level account of the mechanisms that would make the end-to-end confidentiality claim true, and it reports no quantitative accuracy evaluation against any ground truth. The paper's significance as a privacy-preserving machine learning demonstration therefore rests on unverified assertions, while its significance as a manufacturing image analysis case study is limited by the absence of performance metrics.

major comments (3)
  1. [§3.3.1 and §3.3.2] The central privacy claim that 'neither the platform itself nor other users can decrypt or access the data at any point in the process' (§3.3.1) is contradicted by the implementation description in §3.3.2. The user's image is transferred via an HTTP POST request to a Flask backend, passed to AWS Lambda, and processed on an Amazon EC2 instance running Python and TensorFlow; no homomorphic encryption, secure enclave, or equivalent mechanism is described in this flow. Unless such a mechanism exists and is specified, the platform and its cloud provider would see the plaintext image, making the end-to-end confidentiality claim unsupported. This is a load-bearing issue because the paper's main contribution is framed as privacy-preserving.
  2. [§5] The claim that the tool is 'more rapid and accurate' than manual counting is not supported by the reported evidence. The time comparison is between a manual count of one ninth of the image area (about 5 minutes) and an automated count of the entire image (about 30 seconds); these are not commensurate quantities, and the stated '10x decrease' does not follow from the numbers as presented. Accuracy is never measured: there is no ground-truth comparison, no precision/recall or count-error statistic, and no quantification of inter-operator variability. The statement in §4 that operators would 'manually count the crystals in the image to verify the algorithm's results' does not report any agreement metric. The accuracy claim in the abstract should either be removed or supported with data.
  3. [§3.2] The deep learning model is described without any quantitative evaluation. No training set size, number of images, train/test split, hyperparameters, or performance metrics (e.g., precision, recall, F1, count error) are reported for the StarDist-based model; Figure 7 shows only example predictions. Additionally, the small cluster removal module uses thresholds ('20 pixels', '5 pixels') described as 'experimental thresholds' while the text states that 'the internal opening was hardly observed,' which undermines the stated assumption behind the module. A sensitivity analysis or a clear justification for these thresholds is needed for the algorithm to be reproducible and for the accuracy claim to be credible.
minor comments (4)
  1. [§3.2] The reference to the U-Net architecture is incorrect: the text says 'Unet architecture [26]', but reference [26] is Girshick et al. (R-CNN); U-Net is reference [25].
  2. [§1] Reference [30] is cited to support the claim that AI has been used to a limited extent for manufacturing, but reference [30] is a computational pathology dataset paper; this citation does not match the claim.
  3. [§1.1] The statement about imaging techniques applied to stirred tank crystallizers cites reference [20], which is the MyDataCan paper; this appears to be the wrong reference (possibly [22]).
  4. [Throughout] Typographical issues: 'Pascal Delpe -Brice' contains an unintended space in the author list; 'Convex full detection' in §3.2 should read 'convex hull detection'; 'Nielson's' in §3.3.2 should be 'Nielsen's'.

Circularity Check

1 steps flagged · score 6.0 of 10

The crystal-counting tool and its image-processing pipeline are self-contained, but the paper's central privacy-preserving claim rests entirely on a same-author citation to the MyDataCan paper, making that load-bearing premise circular.

  1. self citation load bearing [Section 3.3.1 (The importance of data confidentiality for manufacturers), supported by reference [20]]
    "All data is encrypted during the entire process –upload, in transit to/from machine learning tools, during analysis, and while it is being stored by MyDataCan–so that neither the platform itself nor other users can decrypt or access the data at any point in the process. ... MyDataCan guarantees data privacy, confidentiality, and security using homomorphic encryption, federated learning analysis, secure multi-party computation, and zero-knowledge proofs, as shown in Figure 8. [20]"

    The paper never proves or independently verifies this guarantee. It attributes it to reference [20] (Sweeney, L., Barker, L. B., & Shank, J., 2025), whose authors include two of the present authors, Latanya Sweeney and Joshua Shank. The deployment description in Section 3.3.2—an HTTP POST of the image to a Flask backend, then to AWS Lambda, then to an EC2 instance running TensorFlow—contains no description of encryption, confidential computing, or a secure enclave, so the paper's own implementation does not establish the claimed end-to-end privacy. The conclusion that the work demonstrates how a privacy-preserving platform, MyDataCan, facilitates the sharing of confidential data therefore reduces to the same-group citation rather than to a demonstrated result.

full rationale

The crystal-counting contribution is not circular: the algorithm is built from a published baseline (StarDist [19]), fine-tuned on the authors' own images, and compared in deployment against manual counting; no fitted parameter is relabeled as a prediction, and no known result is merely renamed. The 10x time-saving and seed-count improvements are performance claims, not derivations from the paper's inputs. The circularity is confined to the privacy-preserving platform claim. MyDataCan's guarantees are asserted in Section 3.3.1 and attributed to the authors' own prior paper [20], which is not machine-checked, code-reproduced here, or externally audited. Moreover, the actual deployment flow described in Section 3.3.2 sends image data via HTTP POST to Flask, Lambda, and EC2, with no described encryption mechanism in that flow, so the paper's implementation does not corroborate the cited guarantee. Because this unverified self-citation carries the paper's central premise, the circularity score is 6. The crystal-counting algorithm itself remains independent content, so the score is not higher.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The central algorithmic contribution rests on two hand-set thresholds and a false premise about crystal morphology. The platform contribution rests on an unverified security model drawn from the authors' own prior publication.

free parameters (2)
  • Small cluster removal thresholds (20 pixels, 5 pixels) = 20 px for clusters with at most one crystal; 5 px for overlapping clusters
    These thresholds were described as 'experimental thresholds' set during development, not derived from first principles or validated on a holdout set. They affect which objects are counted as crystals.
  • Image preprocessing brightness reduction = Brightest 20% of colors dimmed by 20%
    This preprocessing step was added to handle background grid interference, but the exact values are choices made by the authors without ablation or sensitivity analysis.
assumptions (2)
  • ad hoc to paper Each crystal has an internal opening (hole) that can be used for detection
    Stated in Section 3.2 as the assumption behind the small cluster removal module. The paper later notes that 'the internal opening was hardly observed', meaning the axiom did not hold for the actual data and required a correction.
  • domain assumption MyDataCan provides the claimed privacy guarantees (homomorphic encryption, federated learning, secure multi-party computation, zero-knowledge proofs)
    Invoked in Section 3.3.1. These capabilities are asserted based on the authors' prior work (reference [20]) without implementation details or independent verification in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Empowering Manufacturers with Privacy-Preserving AI Tools: A Case Study in Privacy-Preserving Machine Learning to Solve Real-World Problems." pith.science (2026). https://pith.science/paper/NHHPL44K

@misc{pith2026250701808,
  author       = {Pith},
  title        = {Pith review of: Empowering Manufacturers with Privacy-Preserving AI Tools: A Case Study in Privacy-Preserving Machine Learning to Solve Real-World Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHHPL44K}},
  note         = {Machine review of arXiv:2507.01808}
}
read the original abstract

Small- and medium-sized manufacturers need innovative data tools but, because of competition and privacy concerns, often do not want to share their proprietary data with researchers who might be interested in helping. This paper introduces a privacy-preserving platform by which manufacturers may safely share their data with researchers through secure methods, so that those researchers then create innovative tools to solve the manufacturers' real-world problems, and then provide tools that execute solutions back onto the platform for others to use with privacy and confidentiality guarantees. We illustrate this problem through a particular use case which addresses an important problem in the large-scale manufacturing of food crystals, which is that quality control relies on image analysis tools. Previous to our research, food crystals in the images were manually counted, which required substantial and time-consuming human efforts, but we have developed and deployed a crystal analysis tool which makes this process both more rapid and accurate. The tool enables automatic characterization of the crystal size distribution and numbers from microscope images while the natural imperfections from the sample preparation are automatically removed; a machine learning model to count high resolution translucent crystals and agglomeration of crystals was also developed to aid in these efforts. The resulting algorithm was then packaged for real-world use on the factory floor via a web-based app secured through the originating privacy-preserving platform, allowing manufacturers to use it while keeping their proprietary data secure. After demonstrating this full process, future directions are also explored.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [20]

    B., & Shank, J

    Sweeney, L., Barker, L. B., & Shank, J. (2025, May 23). Trustworthy AI Data Communities (AI-DCs): Unlocking the benefits of AI for all, using U.S. manufacturing as a case study. Technology Science. https://techscience.org/a/2025052301/

    work page arXiv 2025

  2. [1]

    Huang, H., Liu, L., & Ngadi, M. (2014). Recent developments in hyperspectral imaging for assessment of food quality and safety. Sensors, 14(4), 7248–7276. https://doi.org/10.3390/s140407248

    work page doi:10.3390/s140407248 2014

  3. [2]

    Bobiak, J., & McGeorge, G. (2015). Assessing hyperspectral image content of pharmaceutical products using the Herfindahl–Hirschman index and Ripley's K- functions. Applied Spectroscopy, 69(8), 955–965. https://doi.org/10.1366/14-07650

    work page doi:10.1366/14-07650 2015

  4. [3]

    Su, Q., Ganesh, S., Moreno, M., Bommireddy, Y ., Reklaitis, G., & Nagy, Z. (2019). A perspective on quality-by-control (QbC) in pharmaceutical continuous manufacturing. Computers & Chemical Engineering, 125, 216–231. https://doi.org/10.1016/j.compchemeng.2019.03.001

    work page doi:10.1016/j.compchemeng.2019.03.001 2019

  5. [4]

    Yang, Y ., Pan, L., Ma, J., Yang, R., Zhu, Y ., Yang, Y ., & Zhang, L. (2020). A high- performance deep learning algorithm for the automated optical inspection of laser welding. Applied Sciences, 10(3), 933. https://doi.org/10.3390/app10030933

    work page doi:10.3390/app10030933 2020

  6. [5]

    Tsai, D., Huang, Y ., & Chiu, W. (2021). Deep learning from imbalanced data for automatic defect detection in multicrystalline solar wafer images. Measurement Science and Technology, 32(12), 124003. https://doi.org/10.1088/1361-6501/ac1fbf

    work page doi:10.1088/1361-6501/ac1fbf 2021

  7. [6]

    Nakkina, T., Vinayaka, M., Masad, A., Mansori, M., & Bukkapatnam, S. (2022). Smart defect identification for manufacturing applications. Surface Topography: Metrology and Properties, 10(3), 035045. https://doi.org/10.1088/2051-672X/ac9247

    work page doi:10.1088/2051-672x/ac9247 2022

  8. [7]

    Hashmi, A., Mali, H., Meena, A., Hashmi, M., & Bokde, N. (2023). Surface characteristics measurement using computer vision: A review. Computer Modeling in Engineering & Sciences, 135(2), 917–1005. https://doi.org/10.32604/cmes.2023.021223

    work page arXiv 2023

Show all 31 references
  1. [8]

    Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1), 62–66. https://doi.org/10.1109/TSMC.1979.4310076

    1979

  2. [9]

    Zhou, P ., Ye, W., & Wang, Q. (2017). An improved Canny algorithm for edge detection. In 2017 IEEE International Conference on Software Engineering and Service Science (ICSESS) (pp. 275–278). IEEE. https://doi.org/10.1109/ICSESS.2017.8342913

    2017

  3. [10]

    Lawson, C., & Hanson, R. (1974). Solving least squares problems. Prentice-Hall. https://doi.org/10.1137/1.9781611971217

    1974 doi

  4. [11]

    Schumacher, A., Englert, A., Susin, J., Marczak, L., & Cardozo, N. (2014). An automated measuring methodology for crystal size in sweetened condensed milk using digital image processing and analysis. Food Analytical Methods, 8(7), 1858–

    2014

  5. [12]

    Wu, W. (2023). Sensor fusion and calibration-based adaptive image analysis procedure for in situ crystal size measurement. Crystal Growth & Design, 23(10), 7076–7089. https://doi.org/10.1021/acs.cgd.3c00273

    2023 doi

  6. [13]

    Xie, Y ., & Ji, Q. (2002). A new efficient ellipse detection method. In Proceedings of the 16th International Conference on Pattern Recognition (ICPR 2002) (Vol. 2, pp. 957– 960). IEEE. https://doi.org/10.1109/ICPR.2002.1048464

    2002 arXiv

  7. [14]

    Jain, A. (1986). Fundamentals of digital image processing. Prentice-Hall

    1986

  8. [15]

    Liang, Q., Shakouri, A., & Allebach, J. (2023). Analysis of food processing crystal images. Electronic Imaging, 2023(15), 196-1–196-6. https://doi.org/10.2352/EI.2023.35.15.COLOR-196

    2023 doi

  9. [16]

    Gao, Z., Wu, Y ., Bao, Y ., Gong, J., Wang, J., & Rohani, S. (2018). Image analysis for in- line measurement of multidimensional size, shape, and polymorphic transformation of L-glutamic acid using deep learning-based image segmentation and classification. Crystal Growth & Desi...

    2018 doi

  10. [17]

    He, K., Gkioxari, G., Dollár, P ., & Girshick, R. (2017). Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision (pp. 2961–2969). https://doi.org/10.48550/arXiv.1703.06870

    2017 doi arXiv

  11. [18]

    Huang, Z., Huang, L., Gong, Y ., Huang, C., & Wang, X. (2019). Mask scoring R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 6409–6418). https://doi.org/10.48550/arXiv.1903.00241

    2019 doi arXiv

  12. [19]

    Schmidt, U., Weigert, M., Broaddus, C., & Myers, G. (2018). Cell detection with star- convex polygons. In A. Frangi, J. Schnabel, C. Davatzikos, C. Alberola-López, & G. Fichtinger (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2018 (pp. 265–273). S...

    2018 doi

  13. [21]

    Nielsen, J. (1994). 10 usability heuristics for user interface design. Nielsen Norman Group. https://www.nngroup.com/articles/ten-usability-heuristics/

    1994

  14. [22]

    Huo, Y ., Liu, T., Liu, H., Cai, Y ., & Wang, X. (2016). In-situ crystal morphology identification using imaging analysis with application to the L-glutamic acid crystallization. Chemical Engineering Science, 148, 126–139. https://doi.org/10.1016/j.ces.2016.03.039

    2016 doi

  15. [23]

    Nagy, Z. K. & Braatz, R. D. (2012). Advances and new directions in crystallization control. Annual Review of Chemical and Biomolecular Engineering, 3(1), 55-75. https://doi.org/10.1146/annurev-chembioeng-062011-081043

    2012 doi

  16. [24]

    Stapley, A., Himawan, C., MacNaughtan, W., & Foster, T. (2009). A computational method for extracting crystallization growth and nucleation rate data from hot stage microscope images. Crystal Growth & Design, 9(12), 5061–5068. https://doi.org/10.1021/cg9000413

    2009 doi

  17. [25]

    Ronneberger, O., Fischer, P ., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. In N. Navab, J. Hornegger, W. M. Wells, & A. F . Frangi (Eds.), Medical image computing and computer-assisted intervention – MICCAI 2015 (Vol. 9351, pp. 234–241)....

    2015 doi

  18. [26]

    Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 580–587). https://doi.org/10.1109/CVPR.2014.81

    2014 doi

  19. [27]

    Wu, X., Meng, Y ., Zhang, J., Wei, J., & Zhai, X. (2023). Amodal segmentation of cane sugar crystal via deep neural networks. Journal of Food Engineering, 348, 111435. https://doi.org/10.1016/j.jfoodeng.2023.111435

    2023

  20. [28]

    Impoco, G., Tuminello, L., Fuca, N., Caccamo, M., & Licitra, G. (2011). Segmentation of structural features in cheese micrographs using pixel statistics. Computers and Electronics in Agriculture, 79(2), 199–206. https://doi.org/10.1016/j.compag.2011.09.013

    2011 doi

  21. [29]

    M., & Whitefoot, K

    Donofrio, N. M., & Whitefoot, K. S. (2015). Making value for America: Embracing the future of manufacturing, technology, and work. National Academy of Engineering of the National Academies. https://doi.org/10.17226/19483

    2015 doi

  22. [30]

    Kumar, N., Verma, R., Sharma, S., Bhargava, S., Vahadane, A., & Sethi, A. (2017). A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging, 36(7), 1550–1560. https://doi.org/10.1109/TMI.2017.2677499

    2017

  23. [1867]

    https://doi.org/10.1007/s12161-014-0054-x

    doi

Pith tools

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