Pith. sign in

REVIEW 5 major objections 6 minor 51 references

Detecting Content Rating Violations in Android Applications: A Vision-Language Approach

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read A vision-language model trained on app icons, screenshots, and descriptions can predict Google Play content ratings and shortlist apps whose declared rating is likely wrong, including nine 'Teacher Approved' games.

desk verdict Plausible VL triage for app content ratings, but the violation-detection claim is built on developer labels and a one-directional audit. read the letter →

arxiv 2502.15739 v1 pith:7ZMURSIN submitted 2025-02-07 cs.LG cs.CVcs.MM

classification cs.LGcs.CVcs.MM
keywords MobileAppsContentRatingse-SafetyAndroidVision-LanguageModelsCLIPstyleencodercontrastivelearning
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

Google Play content ratings are self-reported by developers, and the store's scale makes manual verification impractical. This paper tries to establish that a vision-language model trained on static app metadata—icons, screenshots, and descriptions—can both predict a game's official content rating and flag apps whose declared rating is probably too low. The authors report about 6% relative accuracy gain over a fine-tuned CLIP model, and their in-the-wild flags point to real-world signal: apps flagged as likely violations were removed from the store at a higher rate than correctly predicted apps (34.5% vs 27.4%). If true, the model gives regulators and app stores a cheap, automatic triage tool that reduces the manual inspection burden.

What carries the argument

Two parallel image encoders: a content encoder built on the CLIP ViT-B/16 image encoder, and a style encoder trained with a masked siamese objective—randomly masked 3x3 patches are fed to an online network whose outputs must match an exponential-moving-average target network's outputs on unmasked images, so the branch learns texture, colour, and artistic style rather than objects. A frozen 110M-parameter RoBERTa text encoder processes randomly chunked sentences from the app description, and a custom cross-attention transformer stack lets every image patch attend over text tokens and vice versa, closing the semantic gap between creatives and descriptions. The final image embedding is the content embedding plus a scaled style embedding (alpha=0.1), and the model is trained with Sigmoid contrastive loss (SigCL) using content-rating labels as supervision, plus an MSE masked-reconstruction term; classification is done by MLP heads on the fused embeddings with majority voting across an app's images.

What would settle it

Take a random sample of apps flagged as potential malpractices and have independent human reviewers assign content ratings from the app's full store listing; if the flagged apps turn out to be no more likely to deserve higher ratings than a matched sample of apps the model did not flag, the central claim fails. A simpler test: re-run the deletion-rate comparison controlling for download count and app age—if the higher deletion rate of flagged apps disappears, the model is not finding policy-violating apps, just unpopular ones.

Watch

Extended reading notes

Core claim

The central claim is that content ratings for mobile games are predictable from static Play Store creatives, and that prediction mismatches expose real rating malpractices. The paper's model separates visual style from visual content—a masked siamese style encoder captures cartoonish, glittery, or dark aesthetics while a CLIP-based content encoder captures objects and scenes—then aligns image patches with words from the app description through cross-attention, and trains the whole thing with a supervised Sigmoid contrastive loss that treats image-text pairs sharing the same rating as positives. On 10,000 test games it reaches 53.1% accuracy across G, PG, M, MA15+, and R18+, a relative improvement of about 5.9% over pre-trained CLIP and 5.8% over CLIP fine-tuned on the same data. When the model predicts a higher (mature) rating than the developer declared, manual review of 350 such apps found 62 (17.7%) that look like genuine violations, and those flagged apps were later removed from the Play Store at a higher rate than correctly predicted apps, which the paper reads as evidence that the flags carry enforcement-relevant signal.

Load-bearing premise

The method assumes that the developer-declared ratings on the Google Play Store, especially for popular games, are mostly correct, so the model learns the true rating signal from them and any mismatch between prediction and declaration can be read as a likely violation.

Editorial extensions

If this is right

  • A regulator or store operator can feed static app metadata through the model and get a shortlist of games whose declared rating is probably too low, shrinking the manual review workload from millions of apps to a few hundred.
  • The deletion-rate gap (34.5% flagged vs 27.4% correct) suggests the shortlist is enriched for apps that later get pulled from the store, giving enforcement value, not just label-fitting.
  • Teacher Approved status does not guarantee correctness: of 2,172 such apps, 10.4% were flagged as potential malpractices and nine were manually confirmed as unsuitable for children.
  • The style encoder matters: ablations show removing it drops accuracy by about 2.4 points, and removing cross-attention hurts the majority G/PG/M classes most.
  • Because the pipeline uses only static metadata (icons, screenshots, descriptions), it can be applied at scale without running or installing the apps.

Reading between the lines

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

  • The 53% accuracy ceiling may partly reflect noisy labels rather than model weakness; training on cleaner labels (e.g., manually reviewed apps or Apple App Store ratings) could make the flagging signal stronger, a direction the paper itself notes.
  • The removal-rate evidence is indirect: removals can be developer-initiated or driven by user complaints, so the flags may be identifying apps that attract complaints rather than apps that violate policy; both are useful but distinct signals.
  • The same content/style decomposition could transfer to other safety-relevant app-store tasks, such as flagging gambling-themed or sexualized imagery in non-game apps, or detecting dark patterns in children's advertising.
  • A direct test of the approach's generality would be training the same architecture on another store's rating scale (e.g., Apple's 4+ to 17+) and checking whether cross-store matched apps yield higher agreement.
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

5 major / 6 minor

Summary. The paper proposes a vision-language model for predicting the content ratings of Android games from static Play Store metadata (app icons, screenshots, and descriptions). The architecture combines a CLIP-based content encoder, a masked style encoder with EMA updates, a frozen RoBERTa text encoder, and a cross-attention module, trained with a supervised Sigmoid contrastive loss plus an MSE style-loss term. Experiments on a crawled dataset of roughly 16,000 popular games report a relative accuracy improvement of about 5.9% over a CLIP fine-tuned baseline, and an analysis of predictions in the wild identifies 71 potential content-rating 'malpractices' and 32 potential 'disguises', including nine Teacher Approved apps. The authors also report that apps flagged for malpractices were removed from the Play Store at a higher rate than correctly predicted apps.

Significance. If validated, the proposed approach would be a practically useful tool for e-safety regulators and app-store operators to prioritize manual review of content ratings, since current enforcement is largely complaint-driven. The paper contributes a new architecture for combining content and style visual features with text in a contrastive vision-language framework, a substantial real-world dataset, and a series of ablations isolating the effects of the style encoder, cross-attention, and loss function. The authors also make a good-faith effort to manually inspect a sample of flagged apps and to use deletion rates as an external signal. However, the central claim that the model 'detects content rating violations' is weakened by the reliance on developer-declared ratings as both training supervision and evaluation ground truth, by the one-directional manual audit (only flagged apps are inspected), and by the absence of statistical significance testing for the headline accuracy and deletion-rate comparisons.

major comments (5)
  1. [Sec. VI-B1 / IV-A] The definition of a 'potential malpractice' is a prediction strictly higher than the developer-declared rating, which means the detection signal is literally the classifier's own disagreement with the label. The manual verification of 350 flagged apps (finding 62, or 17.7%, as potential malpractices) provides some evidence of precision, but no non-flagged apps are audited, so the paper never establishes enrichment over the base rate of violations. Without a random audit of both flagged and non-flagged apps with identical verification criteria, the claims in the abstract and Sec. VII that the model 'detected' 71 violations and 32 disguises are not supported.
  2. [Sec. IV-A / Tab. I] The test set is not simply the 'next 10k games': the authors state that due to scarcity they 'expanded our search space' for MA15+ and R18+ and appended such apps to the train, validation, and test sets. This means the test distribution is altered by construction, and the counts reported in Sec. VI (71 malpractices, 32 disguises) and the deletion-rate denominators in Sec. VI-B5 may be driven by the oversampled rare classes. The paper should report results separately for the unaltered top-10k split and for the oversampled test set, and should describe the oversampling procedure precisely (how many apps were added, how they were selected, and whether duplicates were removed).
  3. [Sec. V-A / Tab. II] The headline accuracy comparison is a single-run point estimate with no standard deviations, confidence intervals, or significance tests. The claimed relative improvement of 5.8% over CLIP-fine-tuned (53.09 vs 50.18) could easily fall within run-to-run variation, especially given the small differences in several macro metrics in the same table. The authors should report results over multiple random seeds and a paired significance test (e.g., McNemar's test over the same test items) before claiming a consistent improvement.
  4. [Sec. VI-B5 / VII] The deletion-rate analysis uses Play Store removal as a proxy for policy enforcement, but the paper itself acknowledges that developers can discontinue apps voluntarily. The reported gap (34.5% vs 27.4% for Teacher Approved apps; 45.7% vs 29.1% in the broader set) is not accompanied by any statistical test or confidence interval, and the two removal causes cannot be distinguished from the crawled data. At minimum, the authors should report uncertainty estimates and discuss how voluntary developer abandonment could bias the comparison; ideally, they should attempt to distinguish removal reasons using archived pages or policy-enforcement announcements.
  5. [Sec. VII / Eq. (3)] The paper acknowledges that the assumption of reliable labels for popular apps 'may introduce noise into the CLIP fine-tuning process,' but this label noise is load-bearing because the SigCL objective in Eq. (3) treats all image-text pairs sharing the same developer-declared rating as positive, and the same noisy labels are used for evaluation. The effect of label noise is never quantified. The limitation section mentions cross-referencing Apple App Store ratings, but the paper does not implement this check; without some independent label audit, the accuracy numbers and violation flags remain entangled with the very labels the method is meant to verify.
minor comments (6)
  1. [Sec. III-D / Eq. (3)] The symbol τ is reused for the EMA decay rate in Eq. (1) and for what appears to be a temperature-like parameter in Eq. (3), but the latter τ is never defined. Please use distinct symbols and define all hyperparameters.
  2. [Sec. III-D / Eq. (3)] The notation P is overloaded: it denotes the set of positive pairs in Eq. (3) but also the macro precision metric in Tab. II. Additionally, the bias term b in Eq. (3) is not defined or tied to the text's claim that 'b alleviates the heavy imbalance.'
  3. [Abstract / Sec. I] The abstract says '~6% better relative accuracy' while the contributions section says '5.9% and 5.8% relative improvements' and Tab. II shows 53.09 vs 50.18 for the multi-modal setting (a relative gain of about 5.8%). These numbers should be made consistent.
  4. [Sec. III-B] The description of randomized text chunking ('four or more consecutive sentences randomly extracted') is underspecified: the paper does not state how many chunks are sampled per description, whether chunks can overlap, or how the max token length of 256 is reconciled with average description lengths of over 2,000 words.
  5. [Sec. IV-C] The majority-voting procedure over multiple images is not fully specified: the paper does not state how ties are resolved, what happens when an app has only an icon and no screenshots, or whether the number of images per app varies.
  6. [Sec. VI-B4] The Teacher Approved analysis reports that 10.4% of 2,172 apps were flagged, with 92% of those being G-to-PG flags, but only nine apps were manually verified. The paper should state explicitly that the 10.4% figure is an automated flag rate, not a verified violation rate, and should report the manual verification criteria in more detail.

Circularity Check

0 steps flagged · score 2.0 of 10

No formal circularity: the supervised classifier evaluation is self-contained, and the violation-detection claim is explicitly hedged and corroborated by manual review and deletion-rate evidence.

full rationale

The paper's core result is a supervised vision-language classifier trained on developer-declared Play Store ratings and evaluated against held-out ratings (Sec. IV-V). This is a standard supervised benchmark, not a circular derivation: the SigCL objective in Eq. 3 fits image-text representations to ratings, and the reported ~53.1% accuracy is compared with independently trained CLIP and CLIP-fine-tuned baselines. The violation-detection analysis in Sec. VI-B1 defines 'potential malpractices' as cases where the model predicts a rating higher than the developer label. That is an explicit screening definition, not a logical reduction used as proof: the paper manually reviews 350 extreme disagreements, finding 62 (17.7%); separately audits 2,172 Teacher Approved apps; and reports a higher nine-month deletion rate for flagged apps than for correctly classified apps (34.5% vs 27.4%, Sec. VI-B4/B5). These checks are external, albeit imperfect, behavioral evidence. The self-citations ([37], [46]) are not load-bearing: they support auxiliary claims about app takedowns and developer discontinuations, and no uniqueness theorem or prior same-author result is invoked to force the method. The label-noise concern raised in Sec. VII is a real validity limitation, but it is an assumption about ground truth noise, not a circular reduction where the model is defined in terms of the outcome it is claimed to predict.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new physical entities are postulated. The paper introduces named categories ('potential malpractices', 'potential disguises', 'unverifiable apps') but these are labels for prediction-error patterns, not independent discoveries.

free parameters (4)
  • alpha (style embedding weight) = 0.1
    Empirically selected in Sec. III-A2 for qi = qc + alpha*qs; central to how style features enter the image embedding.
  • lambda (loss trade-off) = 5
    Grid-searched coefficient in L = LSigCL + lambda*Lmse, Sec. IV-B; affects balance between contrastive and style-reconstruction losses.
  • masking hyperparameters = 3 patches of 3x3
    Sec. III-A1 randomly masks three 3x3 patches for the style encoder; chosen without analysis of sensitivity.
  • EMA target decay rate tau
    Sec. III-A1 defines tau in [0,1] for the target network update but does not report the value or schedule used.
assumptions (4)
  • domain assumption Developer-declared Google Play ratings are reliable enough to serve as ground truth.
    Sec. IV-A assumes top apps are well monitored and less noisy; Sec. VII acknowledges label noise as a limitation. If labels are wrong, the model learns a distorted mapping.
  • domain assumption App removal from Google Play is a proxy for content policy violation.
    Sec. VI-B5 uses deletion rates as evidence of enforcement; removal can also reflect developer discontinuation, so the proxy is confounded.
  • domain assumption Static metadata (icons, screenshots, descriptions) contains enough signal to infer content rating.
    The model only sees text and images; Sec. VI-B3 shows some apps carry no visual or textual cues, so this assumption fails for a subset of apps.
  • domain assumption CLIP and RoBERTa representations transfer to the app-store domain.
    Sec. III uses a CLIP image encoder and frozen RoBERTa text encoder without demonstrating that their pretraining distributions cover app icons, screenshots, or Play descriptions; the style encoder is added partly to compensate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detecting Content Rating Violations in Android Applications: A Vision-Language Approach." pith.science (2026). https://pith.science/paper/7ZMURSIN

@misc{pith2026250215739,
  author       = {Pith},
  title        = {Pith review of: Detecting Content Rating Violations in Android Applications: A Vision-Language Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7ZMURSIN}},
  note         = {Machine review of arXiv:2502.15739}
}
read the original abstract

Despite regulatory efforts to establish reliable content-rating guidelines for mobile apps, the process of assigning content ratings in the Google Play Store remains self-regulated by the app developers. There is no straightforward method of verifying developer-assigned content ratings manually due to the overwhelming scale or automatically due to the challenging problem of interpreting textual and visual data and correlating them with content ratings. We propose and evaluate a visionlanguage approach to predict the content ratings of mobile game applications and detect content rating violations, using a dataset of metadata of popular Android games. Our method achieves ~6% better relative accuracy compared to the state-of-the-art CLIP-fine-tuned model in a multi-modal setting. Applying our classifier in the wild, we detected more than 70 possible cases of content rating violations, including nine instances with the 'Teacher Approved' badge. Additionally, our findings indicate that 34.5% of the apps identified by our classifier as violating content ratings were removed from the Play Store. In contrast, the removal rate for correctly classified apps was only 27%. This discrepancy highlights the practical effectiveness of our classifier in identifying apps that are likely to be removed based on user complaints.

Figures

Figures reproduced from arXiv: 2502.15739 by the authors.

Figure 1
Figure 1. (a): Vision-language model architecture during the training stage. (b): Custom transformer block with cross-attention. (c) Pipeline for the downstream [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Disparity between the content and style of app icons and screenshots. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Examples belonging to 1) potential malpractices, and 2) potential disguises. For each app, the image on the left represents the app icon, and on the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Confusion matrix comparing our method against baselines - using [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualisation of image patches attending to text tokens in the custom [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Examples of unverifiable apps with developer defined content rating [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Examples of teacher-approved apps with incorrect content ratings - A [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: App deletion rates w.r.t number of downloads. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 32 canonical work pages

  1. [1]

    Number of smartphone users worldwide (bil- lions),

    A. Turner, “Number of smartphone users worldwide (bil- lions),” 2024. [Online]. Available: https://www.bankmycell.com/blog/ how-many-phones-are-in-the-world#part-1

  2. [2]

    Identifying and analyzing the privacy of apps for kids,

    M. Liu, H. Wang, Y . Guo, and J. Hong, “Identifying and analyzing the privacy of apps for kids,” in Proceedings of the 17th International Workshop on Mobile Computing Systems and Applications , 2016, pp. 105–110

  3. [3]

    The common sense census: Media use by tweens and teens,,

    “The common sense census: Media use by tweens and teens,,” 2019. [Online]. Available: https://www.commonsensemedia.org/research/ the-common-sense-census-media-use-by-tweens-and-teens-2019

  4. [4]

    Build teacher approved apps,

    “Build teacher approved apps,” https://play.google.com/console/about/ programs/teacherapproved/, accessed: 2024-08-07

  5. [5]

    Automatic content inspection and forensics for children Android apps,

    Q. Luo, J. Liu, J. Wang, Y . Tan, Y . Cao, and N. Kato, “Automatic content inspection and forensics for children Android apps,” IEEE Internet of Things Journal, vol. 7, no. 8, pp. 7123–7134, 2020

  6. [6]

    Qustodio releases 2023 annual report, born connected: The rise of the ai generation,

    Qustodio, “Qustodio releases 2023 annual report, born connected: The rise of the ai generation,” 2023. [Online]. Avail- able: https://static.qustodio.com/public-site/uploads/2024/01/19122535/ ADR 2023-24 EN.pdf

  7. [7]

    Not seen, not heard in the digital world! measuring privacy practices in children’s apps,

    R. Sun, M. Xue, G. Tyson, S. Wang, S. Camtepe, and S. Nepal, “Not seen, not heard in the digital world! measuring privacy practices in children’s apps,” in Proceedings of the ACM Web Conference 2023 , 2023, pp. 2166–2177

  8. [8]

    Developer content policy,

    “Developer content policy,” https://support.google.com/googleplay/ android-developer/answer/9859655?sjid=12738238277494853083-AP& visit id=638689148734509737-3391379066&rd=1#questionnaire, accessed: 2024-12-05

Show all 51 references
  1. [9]

    Android’s latest statistics 2024: How many people have Androids?

    Sam Nguyen, “Android’s latest statistics 2024: How many people have Androids?” September 06, 2023. [Online]. Available: https: //avada.io/articles/how-many-people-have-androids/

  2. [10]

    Developer content policy,

    “Developer content policy,” https://developer.apple.com/app-store/ review/guidelines/, accessed: 2024-12-05

  3. [11]

    Mobile apps for kids: Current privacy disclosures are disappointing,

    F. T. Commission, “Mobile apps for kids: Current privacy disclosures are disappointing,” Washington, DC. RetrievedAugust, vol. 21, p. 2022, 2012

  4. [12]

    Freely given consent? studying consent notice of third-party tracking and its violations of gdpr in Android apps,

    T. T. Nguyen, M. Backes, and B. Stock, “Freely given consent? studying consent notice of third-party tracking and its violations of gdpr in Android apps,” in Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security , 2022, pp. 2369–2383

  5. [13]

    Measuring user perception for detecting unexpected access to sensitive resource in mobile apps,

    T. T. Nguyen, D. C. Nguyen, M. Schilling, G. Wang, and M. Backes, “Measuring user perception for detecting unexpected access to sensitive resource in mobile apps,” in Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security , 2021, pp. 578– 592

  6. [14]

    Protecting your children from inappropriate content in mobile apps: An automatic ma- turity rating framework,

    B. Hu, B. Liu, N. Z. Gong, D. Kong, and H. Jin, “Protecting your children from inappropriate content in mobile apps: An automatic ma- turity rating framework,” in Proceedings of the 24th ACM International on Conference on Information and Knowledge Management , 2015, pp. 1111–1120

  7. [15]

    Can apps play by the coppa rules?

    I. Liccardi, M. Bulger, H. Abelson, D. J. Weitzner, and W. Mackay, “Can apps play by the coppa rules?” in 2014 Twelfth Annual International Conference on Privacy, Security and Trust . IEEE, 2014, pp. 1–9

  8. [16]

    Is this app safe for children? a comparison study of maturity ratings on Android and iOS applications,

    Y . Chen, H. Xu, Y . Zhou, and S. Zhu, “Is this app safe for children? a comparison study of maturity ratings on Android and iOS applications,” in Proceedings of the 22nd international conference on World Wide Web, 2013, pp. 201–212

  9. [17]

    Automatic maturity rating for Android apps,

    C. Zhou, X. Zhan, L. Li, and Y . Liu, “Automatic maturity rating for Android apps,” in Proceedings of the 13th Asia-Pacific Symposium on Internetware, 2022, pp. 16–27

  10. [18]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning . PMLR, 2020, pp. 1597–1607

  11. [19]

    Barlow twins: Self-supervised learning via redundancy reduction,

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Barlow twins: Self-supervised learning via redundancy reduction,” in International conference on machine learning . PMLR, 2021, pp. 12 310–12 320

  12. [20]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

  13. [21]

    Scaling up visual and vision-language representation learning with noisy text supervision,

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig, “Scaling up visual and vision-language representation learning with noisy text supervision,” in International conference on machine learning . PMLR, 2021, pp. 4904–4916

  14. [22]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748 , 2018

  15. [23]

    Lit: Zero-shot transfer with locked-image text tuning,

    X. Zhai, X. Wang, B. Mustafa, A. Steiner, D. Keysers, A. Kolesnikov, and L. Beyer, “Lit: Zero-shot transfer with locked-image text tuning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 18 123–18 133

  16. [24]

    Clip2video: Mastering video- text retrieval via image clip,

    H. Fang, P. Xiong, L. Xu, and Y . Chen, “Clip2video: Mastering video- text retrieval via image clip,” arXiv preprint arXiv:2106.11097 , 2021

  17. [25]

    A straightforward framework for video retrieval using clip,

    J. A. Portillo-Quintero, J. C. Ortiz-Bayliss, and H. Terashima-Mar ´ın, “A straightforward framework for video retrieval using clip,” in Mexican Conference on Pattern Recognition . Springer, 2021, pp. 3–12

  18. [26]

    Glide: Towards photorealistic image gen- eration and editing with text-guided diffusion models,

    A. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. McGrew, I. Sutskever, and M. Chen, “Glide: Towards photorealistic image gen- eration and editing with text-guided diffusion models,” arXiv preprint arXiv:2112.10741, 2021

  19. [27]

    Explaining clip’s performance disparities on data from blind/low vision users,

    D. Massiceti, C. Longden, A. Slowik, S. Wills, M. Grayson, and C. Morrison, “Explaining clip’s performance disparities on data from blind/low vision users,” arXiv preprint arXiv:2311.17315 , 2023

  20. [28]

    Evaluating clip: towards characterization of broader capabilities and downstream implications,

    S. Agarwal, G. Krueger, J. Clark, A. Radford, J. W. Kim, and M. Brundage, “Evaluating clip: towards characterization of broader capabilities and downstream implications,” arXiv preprint arXiv:2108.02818, 2021

  21. [29]

    Stable bias: Evaluating societal representations in diffusion models,

    S. Luccioni, C. Akiki, M. Mitchell, and Y . Jernite, “Stable bias: Evaluating societal representations in diffusion models,” Advances in Neural Information Processing Systems , vol. 36, 2024

  22. [30]

    Masked siamese networks for label-efficient learning,

    M. Assran, M. Caron, I. Misra, P. Bojanowski, F. Bordes, P. Vincent, A. Joulin, M. Rabbat, and N. Ballas, “Masked siamese networks for label-efficient learning,” in European Conference on Computer Vision . Springer, 2022, pp. 456–473

  23. [31]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 000–16 009

  24. [32]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 . Springer,...

  25. [33]

    From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions,

    P. Young, A. Lai, M. Hodosh, and J. Hockenmaier, “From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions,” Transactions of the Association for Computational Linguistics, vol. 2, pp. 67–78, 2014

  26. [34]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

  27. [35]

    Sigmoid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sigmoid loss for language image pre-training,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 11 975–11 986

  28. [36]

    Unified contrastive learning in image-text-label space,

    J. Yang, C. Li, P. Zhang, B. Xiao, C. Liu, L. Yuan, and J. Gao, “Unified contrastive learning in image-text-label space,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 19 163–19 173

  29. [37]

    Early detection of spam mobile apps,

    S. Seneviratne, A. Seneviratne, M. A. Kaafar, A. Mahanti, and P. Mo- hapatra, “Early detection of spam mobile apps,” in Proceedings of the 24th International Conference on World Wide Web , 2015, pp. 949–959

  30. [38]

    A multi-modal neural embeddings approach for detecting mobile counterfeit apps,

    J. Rajasegaran, N. Karunanayake, A. Gunathillake, S. Seneviratne, and G. Jourjon, “A multi-modal neural embeddings approach for detecting mobile counterfeit apps,” in The World Wide Web Conference, 2019, pp. 3165–3171

  31. [39]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014. 11

  32. [40]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” arXiv preprint arXiv:1608.03983 , 2016

  33. [41]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

  34. [42]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

  35. [43]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” in International conference on machine learning . PMLR, 2022, pp. 12 888–12 900

  36. [44]

    A multiscale visualization of attention in the transformer model,

    J. Vig, “A multiscale visualization of attention in the transformer model,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations . Florence, Italy: Association for Computational Linguistics, Jul. 2019, pp. 37–42. [On...

  37. [45]

    Manage target audience and app content settings

    G. P. C. Help, “Manage target audience and app content settings.” [Online]. Available: https://support.google.com/googleplay/ android-developer/answer/9867159?hl=en

  38. [46]

    Detecting and characterising mobile app metamorphosis in Google Play store,

    D. Denipitiyage, B. Silva, K. Gunathilaka, S. Seneviratne, A. Mahanti, A. Seneviratne, and S. Chawla, “Detecting and characterising mobile app metamorphosis in Google Play store,” arXiv preprint arXiv:2407.14565, 2024

  39. [47]

    How we fought bad apps and bad actors in 2023,

    “How we fought bad apps and bad actors in 2023,” https://security.googleblog.com/2024/04/ how-we-fought-bad-apps-and-bad-actors-in-2023.html, 2024, accessed: 2024-08-07

  40. [48]

    A multi-modal neural embeddings approach for detecting mobile counterfeit apps: A case study on Google Play store,

    N. Karunanayake, J. Rajasegaran, A. Gunathillake, S. Seneviratne, and G. Jourjon, “A multi-modal neural embeddings approach for detecting mobile counterfeit apps: A case study on Google Play store,” IEEE Transactions on Mobile Computing , vol. 21, no. 1, pp. 16–30, 2020

  41. [49]

    Comparing apples to Androids: Discovery, retrieval, and matching of and Android apps for cross-platform analyses,

    M. Steinb ¨ock, J. Bleier, M. Rainer, T. Urban, C. Utz, and M. Lindorfer, “Comparing apples to Androids: Discovery, retrieval, and matching of and Android apps for cross-platform analyses,” in 2024 IEEE/ACM 21st International Conference on Mining Software Repositories (MSR) . ...

  42. [2012]

    He received his PhD from the University of Tennessee (USA) in 1995. His research is in data mining and machine learning with a specialization in spatio-temporal data mining, outlier detection, class imbalanced classification, and adversarial learning

  43. [2015]

    Before moving into research, he worked nearly six years in the telecommunications industry in core network plan- ning and operations

    His current research interests include privacy and security in mobile systems, AI applications in security, and behavior biometrics. Before moving into research, he worked nearly six years in the telecommunications industry in core network plan- ning and operations. He receive...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.