Analysing User Reviews to Identify User Concerns Around Permissions in AI Apps

Babar Shah; Faheem Ullah; Muhammad Junaid; Muhammad Moiz Khalid; Myles Watkinson; Tehmina Karamat Khan

REVIEW 3 major objections 4 minor 29 references

This paper shows that permission-related app reviews can be classified from the full review stream with 82% accuracy using AI-generated examples as training labels, and that those reviews cluster by user sentiment rather than by permission

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 08:57 UTC pith:G62FYE5G

load-bearing objection The GPT-4-generated-review idea is worth discussing, but the 82% accuracy overstates what the evaluation supports, and the sentiment clusters likely inherit generator bias. the 3 major comments →

arxiv 2607.29343 v1 pith:G62FYE5G submitted 2026-07-31 cs.LG

Analysing User Reviews to Identify User Concerns Around Permissions in AI Apps

Babar Shah , Faheem Ullah , Myles Watkinson , Muhammad Moiz Khalid , Tehmina Karamat Khan , Muhammad Junaid This is my paper

classification cs.LG

keywords app reviewspermissionsAI mobile appsreview classificationsynthetic training datasentiment clusteringprivacy concernsmachine learning

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that permission-related concerns can be pulled out of app-store review streams without a manually labeled training set. The method uses a large language model to generate a small set of prototype permission and security reviews, then finds human reviews that closely resemble those prototypes to serve as training labels. A binary classifier trained on those selected reviews identifies permission-related reviews from the full review dataset with 0.82 accuracy. Clustering the identified reviews shows that users' concerns group by sentiment toward the app—hostility, questioning, disappointment, warning—rather than by the specific permission type involved. If correct, this gives developers and app stores a low-cost way to monitor permission-related backlash at scale.

Core claim

The paper's central claim is that permission-related reviews can be meaningfully classified from a dataset spanning all review types, with accuracy 0.82, and that the topics users raise in those reviews are organized by sentiment rather than by permission type. On the paper's own pipeline, a privacy/security classifier first isolates 20,622 security reviews from 16.1 million human reviews; a second classifier then distinguishes permission-focused reviews from among them (0.71 from security reviews, 0.82 from all reviews) and selects 6,532 permission-related reviews. Clustering those reviews with k=7 produces topics labeled Hostile, Question, Statement, Review, Disappointment, Warning, and Re

What carries the argument

The load-bearing mechanism is prototype-based training-label selection. A large language model generates a small set of permission and security review prototypes; a sentence-embedding model converts both prototypes and the human review corpus into vectors; any human review whose similarity to a prototype exceeds an 80% threshold is treated as a training example for that category. Those automatically selected examples train a feed-forward binary classifier with a pre-trained embedding layer. The same embedding representation feeds a K-means clustering step (k=7 chosen by silhouette score) to produce the sentiment-defined topics. The 80% threshold and the prototype set together define what cou

Load-bearing premise

The load-bearing premise is that a human review with high embedding similarity to an AI-generated permission prototype really is a permission review; if that mapping is systematically biased, the 0.82 accuracy and the sentiment-based topic clusters inherit the bias.

What would settle it

Build a test set of at least 1,000 permission and non-permission reviews with multiple independent human annotators, sampled across app genres, and measure the classifier's accuracy per genre. If overall accuracy drops below roughly 0.75 on this larger set, or if any genre is near chance, the 0.82 claim does not generalize. A cheaper test: generate training labels with two different language models (or two similarity thresholds) and measure inter-label agreement; low agreement would show the labels are an artifact of the generator rather than of the reviews.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

App stores can surface permission-related reviews automatically, because the classifier operates on the full review stream without hand-labeled training data.
Developers can see which permission requests generate hostility and can address those requests directly, since topic clusters reflect emotional response rather than a list of permission names.
Genre differences matter: finance apps attract larger shares of hostile, questioning, and disappointed reviews, while communication apps concentrate in the 'Review' cluster, implying users' expectations about data use vary by app purpose.
Warning and Removal clusters are small, suggesting that outright perceived misuse is rare; the larger clusters point to frustration and confusion as the main permission-related problem.
The pipeline is transferable: generating prototypes for other policy topics (e.g., data sharing, ads, account deletion) would let the same approach monitor additional user concerns.

Where Pith is reading between the lines

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

Inference: If sentiment-based clustering is real, permission-prompt design should be genre-specific—finance apps should preemptively explain why sensitive data is requested, because hostile reviews are concentrated there.
Inference: The 80% similarity threshold and the choice of generator jointly define the target concept; comparing labels from a different generator or a different threshold would quantify how much of the 0.82 accuracy is inherited from that choice.
Inference: The clusters read like an escalation ladder (Question → Disappointment → Hostile → Removal), which could be turned into a single risk score per app update rather than a list of topics.
Inference: The same pipeline could serve as a regression-monitoring signal: a sudden shift in the ratio of Hostile to Question reviews after an update would flag permission changes that upset users before they uninstall.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

The GPT-4-generated-review idea is worth discussing, but the 82% accuracy overstates what the evaluation supports, and the sentiment clusters likely inherit generator bias.

read the letter

Two things to know: the paper has a fresh idea—use GPT-4-generated permission reviews as semantic templates to pull training examples from real app reviews—and that idea is worth airing. But the reported 82% accuracy is not trustworthy as it stands, and the sentiment-based topic clusters are probably contaminated by the generation process.

What the paper does well: the pipeline is clearly described, the authors are explicit about the two stages (security first, then permission), and they don't hide the 71% accuracy for the harder two-stage task. The observation that clusters organize by sentiment rather than permission type is genuinely interesting and worth testing with cleaner data. They also engage with the relevant literature and flag the bias risk in the limitations section.

Where it falls down: the headline accuracy comes from the easier single-stage task on a 100-review balanced test set that appears to be selected by the same GPT-4 similarity rule used for training, so it may only show the model can imitate the prototypes. The paper acknowledges this ('all selected data was biased to the generated data') but does nothing to rule it out, e.g., by testing different generators or thresholds. The 'best version across 20 epochs used for testing' is consistent with optimistic epoch selection. No code, data, or prompts are released. The abstract claims manual annotation is eliminated, but they manually annotated test sets; the 30% vs 50% alignment with Nema et al. is not 'nearly aligning'; and the text calls recall 'moderate' when Table 1 shows 0.87. None of these kill the core hypothesis, but together they mean the paper needs serious revision.

Who should read it: people working on weakly-supervised text classification and anyone thinking about using LLM-generated labels in empirical software engineering. It deserves peer review, but a referee should ask for independent human labels, confidence intervals, an n-gram baseline, and released artifacts.

Referee Report

3 major / 4 minor

Summary. The paper proposes a two-stage pipeline for identifying permission-related reviews in Google Play reviews of AI apps. It scrapes a large human-review corpus, uses GPT-4 generated permission/security prototypes with Universal Sentence Encoder embeddings to select training positives by cosine similarity, trains binary classifiers (security first, then permission), and then applies k-means clustering to the classified permission reviews. The paper reports 0.82 accuracy for permission classification from all reviews and concludes that user concerns cluster by sentiment rather than by permission type.

Significance. If the method's claims hold, the contribution is practically interesting: it would allow app-store stakeholders to screen permission-related reviews without manual annotation, and the sentiment-based topic finding could inform permission-request design. The idea of bootstrapping labels from AI-generated prototypes to overcome data scarcity is creative, and the two-stage architecture is clearly described. However, the central claims currently rest on an unvalidated generator-based labeling step and on small or ambiguously described test sets, so the significance is contingent on additional validation.

major comments (3)

[Section 3.2, Section 4.1] Label validity and circularity: positive training examples for the permission classifier are human reviews with >0.8 cosine similarity to GPT-4 permission-review prototypes, and the RQ2 corpus is selected by the same rule. The classifier may therefore learn to detect 'similar to GPT-4's notion of a permission review' rather than 'a user expressing a permission concern.' The paper concedes this in Section 6 but does not test it. This is load-bearing: the 0.82 accuracy does not establish performance against human judgments. Concretely, the authors should have independent human annotators label a random sample of reviews from the full 16M corpus (not pre-filtered by GR similarity), compare classifier predictions to those labels, and report precision/recall plus inter-annotator agreement. They should also vary the generator (e.g., GPT-3.5, Llama, or no generator) and the similarity threshold
[Section 3.3, Table 1] Test-set construction and epoch selection: the text states 'the best version across 20 epochs used for testing.' Selecting the best epoch based on the test set is a form of test-set leakage and will inflate the reported accuracy (0.82). Moreover, the test set for the permission classifier is not clearly specified: Section 3.2 describes a 100-review manually annotated set for security, then says the process was repeated, but Table 1 gives no test-set size or composition for the permission rows. The authors should use a separate validation set for model selection, a fresh human-labeled test set for final evaluation, and report confidence intervals (e.g., Wilson interval) for accuracy and F1.
[Section 4.2, Table 2] Cluster validity: the conclusion that topics are 'shaped by user sentiment, not by the specific type of permission involved' is based on seven k-means clusters manually labeled with sentiment names. No quantitative validation is reported (e.g., cluster purity against permission type, inter-rater agreement on the topic labels), and silhouette score only measures geometric separation. Since the clustered corpus was selected by GPT-4 prototype similarity, the apparent sentiment structure may reflect the prototypes rather than actual user concerns. The authors should test whether sentiment clusters are reproducible on an independently human-labeled sample, and should compare cluster membership against permission type as an alternative explanation.

minor comments (4)

[Section 3.3] Typo in heading: 'T raining and testing privacy classifier' should be 'Training and testing privacy classifier.'
[Section 3.1] The phrase 'any HR correlating above 80%' is imprecise; the method uses cosine similarity, not correlation. Please use 'cosine similarity exceeding 0.8.'
[Table 1 caption] Please state the number and composition of test reviews for each row. Currently the reader cannot tell whether the 0.82 row is evaluated on a human-annotated set or on an automatically selected set.
[Figure 2] The caption 'Similarity matrix values' is vague. Clarify that subfigure (a) shows box plots of pairwise similarity among generated reviews, and (b) shows silhouette scores from a single k-means run; no error bars or multiple runs are shown.

Circularity Check

0 steps flagged

No circularity: accuracy and sentiment-cluster findings are empirical measurements, not reductions to the GPT-based selection filter.

full rationale

The paper's central claim is an empirical measurement rather than a derivation from its own inputs. The 80% cosine-similarity threshold to GPT-4 permission prototypes defines how positive training examples and the analysis corpus are selected (Section 3.2: "any HR correlating above 80% treated as relating to the same topic"), and the paper concedes in Section 6 that "it could be argued that all selected data was biased to the generated data." However, the reported 0.82 accuracy is not logically forced by that threshold: the same section describes a "manually annotated test set" and the permission classification process is repeated with the same manual annotation, providing an independent (if small) ground-truth check. The sentiment-cluster conclusion in RQ2 comes from k-means clustering of the extracted reviews and manual reading of clusters, not from equating the clusters with the GPT prototype labels; the clustering could in principle have produced permission-type clusters. The self-citations in the reference list are background literature and are not load-bearing for the accuracy or clustering findings. The GPT-filter bias is a real external-validity threat, but it is not a circular definitional reduction of the claimed results to the input labels, and the paper explicitly acknowledges this limitation. No specific reduction by construction can be exhibited, so no circularity step is identified.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper adds no invented entities. Its free parameters concentrate in the label-generation pipeline: the 0.80 similarity threshold (which defines both the training set and the 6,532-review analysis corpus), the GR diversity gates (median similarity 0.5/0.6), the balanced 100-review test sets, and k=7. The load-bearing axiom is that GPT-4 permission-review prototypes are a valid, unbiased proxy for human permission concerns; the authors flag this risk in Section 6 but provide no sensitivity analysis. Secondary domain assumptions: USE cosine similarity >0.8 equals topical equivalence; the scraped Li et al. corpus represents AI-app users; reviews faithfully record concerns; and the medical-imaging '10x data' heuristic transfers to text classification.

free parameters (5)

HR-GR cosine similarity threshold = 0.80
Labels an HR as a permission/security review if its USE embedding correlates >0.80 with any GPT-4 prototype (Section 3.2); it determines the training sets and the 6,532-review RQ2 corpus and is neither swept nor justified against a human-labeled calibration set.
Cluster count k = 7
Selected via silhouette maximum across k=2..10 (Figure 2b); data-driven, but the RQ2 topic structure and genre comparisons depend on this choice.
Test-set class balance and size = 50 positive / 50 negative, n=100 per classifier
Reported accuracies (0.79, 0.71, 0.82) are measured on artificially balanced 100-review sets with no confidence intervals; with a corpus base rate of ~0.04% permission reviews, these numbers do not describe operating performance (Sections 3.2-3.3).
GR diversity gates (median pairwise similarity) = 0.5 (security), 0.6 (permission)
Ad hoc acceptance criterion asserting the generated sets are 'diverse enough' (Section 3.1, Figure 2a); no sensitivity analysis.
Epoch selection = best of 20
The 'best version across 20 epochs' is used for testing (Section 3.3); if chosen on the test set, the reported accuracies are optimistically biased.

axioms (6)

ad hoc to paper GPT-4-generated permission reviews are a valid and unbiased proxy for the content and language of human permission-related reviews.
All training labels and the clustering corpus are defined by >80% similarity to these 100 prototypes (Sections 3.1-3.2); Section 6 concedes 'all selected data was biased to the generated data' as a possible objection, and the 100-review test set is too small to detect systematic bias.
domain assumption Cosine similarity >0.8 in USE embedding space corresponds to topical equivalence for app reviews.
Used to convert GRs into HR labels (Section 3.2); no calibration of the threshold against human judgment, and the matching rule (per-GR max vs. average) is unspecified.
domain assumption The 16.1M scraped reviews from Li et al.'s 56,682 AI-app list are representative of AI-app user reviews.
Data curation (Section 3.1); scraping coverage, language distribution, and platform biases are not discussed.
domain assumption User reviews faithfully reflect genuine user security and privacy concerns.
The motivating premise (Introduction); reviews can be astroturfed, incentivized, or unrepresentative of silent users, and no checks are performed.
standard math Euclidean k-means on USE embeddings produces semantically meaningful topic clusters.
Standard practice, but cluster coherence is only checked by the authors' manual reading after clustering (Section 3.4); no inter-rater reliability or stability analysis.
domain assumption The '10x data rule' for neural network training (5,120 examples for a 512-dim input) applies to this text-classification setting.
Borrowed from medical imaging (Cho et al. [8], Section 3.2) without validation for embeddings and classification heads.

pith-pipeline@v1.3.0-daily-deepseek · 6546 in / 24535 out tokens · 223641 ms · 2026-08-03T08:57:27.409555+00:00 · methodology

0 comments

read the original abstract

Artificial intelligence is increasingly embedded in everyday software, making its integration into mobile apps inevitable. However, AI mobile app developers are not always versed in security and privacy best practices, leaving users to monitor their own security and understand how apps use their data. App reviews capture real user experiences, helping others make informed decisions before downloading. This paper presents a machine learning model for classifying AI app reviews into permission-related categories. Because user reviews are unstructured, assembling a conventional labeled training set is difficult. To address this, AI-generated security and permission reviews are used to identify relevant training examples from a large corpus of human-written reviews, eliminating the need for manual annotation. The proposed approach classified permission reviews with an accuracy of 82%. Analysis shows that users organise their concerns by sentiment toward the requesting app rather than specific permission types, with implications for users, developers, and platform administrators.

Figures

Figures reproduced from arXiv: 2607.29343 by Babar Shah, Faheem Ullah, Muhammad Junaid, Muhammad Moiz Khalid, Myles Watkinson, Tehmina Karamat Khan.

**Figure 2.** Figure 2: Similarity matrix box plots (left) and silhouette score across k values [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 8 canonical work pages

[1]

App store ratings and reviews (2019),https://developer.apple.com/app-store /ratings-and-reviews/

2019
[2]

ACM Comput

Ahmad, H., Dharmadasa, I., Ullah, F., Babar, M.A.: A review on c3i systems’ security: Vulnerabilities, attacks, and countermeasures. ACM Comput. Surv.55(9) (Jan 2023).https://doi.org/10.1145/3558001,https://doi.org/10.1145/35 58001

work page doi:10.1145/3558001 2023
[3]

AI and Society37, 1715–1728 (2021).https://doi.org/10.1007/s00146-021-0 1262-5

Andreotta, A.J., Kirkham, N., Rizzi, M.: Ai, big data, and the future of consent. AI and Society37, 1715–1728 (2021).https://doi.org/10.1007/s00146-021-0 1262-5

work page doi:10.1007/s00146-021-0 2021
[4]

apple.com/en-gb/guide/iphone/iph251e92810/ios

Apple: Control access to information in apps on iphone (nd),https://support. apple.com/en-gb/guide/iphone/iph251e92810/ios
[5]

arXiv preprint arXiv:1803.11175 (2018)

Cer, D., Yang, Y., Kong, S.y., Hua, N., Limtiaco, N., St John, R., Constant, N., Guajardo-Cespedes, M., Yuan, S., Tar, C., Sung, Y.H., Strope, B., Kurzweil, R.: Universal sentence encoder. arXiv preprint arXiv:1803.11175 (2018)

Pith/arXiv arXiv 2018
[6]

International Journal of Innovative Research in Advanced Engineering3, 62–72 (2016).https://doi.org/ 10.6084/M9.FIGSHARE.3504380.V1

Chatterjee, S., Paul, K., Roy, R., Nath, A.: A comprehensive study on security issues in android mobile phone — scope and challenges. International Journal of Innovative Research in Advanced Engineering3, 62–72 (2016).https://doi.org/ 10.6084/M9.FIGSHARE.3504380.V1

work page doi:10.6084/m9.figshare.3504380.v1 2016
[7]

In: Proceedings of ICSE

Chen, N., Lin, J., Hoi, S.C.H., Xiao, X., Zhang, B.: Arminer: mining informative reviews for developers from mobile app marketplace. In: Proceedings of ICSE. pp. 767–778 (2014)

2014
[8]

Cho, J., Lee, K., Shin, E., Choy, G., Do, S.: How much data is needed to train a medical image deep learning system to achieve necessary high accuracy? arXiv preprint arXiv:1511.06348 (2016)

Pith/arXiv arXiv 2016
[9]

arXiv preprint arXiv:1810.04805 (2019)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirec- tional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2019)

Pith/arXiv arXiv 2019
[10]

Eason: Google play store apps reviews scraping and text analytics (sentiment anal- ysis) (2021),https://medium.com/analyticsvidhya/google-play-store-app s-reviews-scraping-and-text-analyticssentiment-analysis-5303294fffa7

2021
[11]

edu/soups/2012/proceedings/a3_Felt.pdf

Felt, A., Ha, E., Egelman, S., Haney, A., Chin, E., Wagner, D.: Android permis- sions: User attention, comprehension, and behavior (nd),https://cups.cs.cmu. edu/soups/2012/proceedings/a3_Felt.pdf

2012
[12]

Grano, G., Di Sorbo, A., Mercaldo, F., Visaggio, C.A., Canfora, G., Panichella, S., Visaggio, C.: Android apps and user feedback: A dataset for software evolution and quality improvement (2017).https://doi.org/10.1145/3121264.3121266

arXiv 2017
[13]

ACM Computing Surveys55(1), 1–36 (2021).https://doi.org/10.1145/3487890

Hu, Y., Kuang, W., Qin, Z., Li, K., Zhang, J., Gao, Y., Li, W., Li, K.: Artificial in- telligence security: Threats and countermeasures. ACM Computing Surveys55(1), 1–36 (2021).https://doi.org/10.1145/3487890

work page doi:10.1145/3487890 2021
[14]

arXiv (2022)

Li, Y., Dang, X., Tian, H., Sun, T., Wang, Z., Ma, L., Klein, J., Bissyande, T.: Ai-driven mobile apps: an explorative study. arXiv (2022)

2022
[15]

arXiv (2021)

Lu, Y., Wang, W., Wei, W., Wang, H.: Machine learning for synthetic data gener- ation: a review. arXiv (2021)

2021
[16]

Mukherjee, D., Ahmadi, A., Pour, M., Reardon, J.: An empirical study on user reviews targeting mobile apps’ security privacy (2020)

2020
[17]

In: Proceedings of the 44th International Conference on Software Engineering (2022).https://doi.org/10.1145/3510003.3510079

Nema, P., Anthonysamy, P., Taft, N., Peddinti, S.T.: Analyzing user perspectives on mobile app privacy at scale. In: Proceedings of the 44th International Conference on Software Engineering (2022).https://doi.org/10.1145/3510003.3510079

arXiv 2022
[18]

Nguyen, D.C., Derr, E., Backes, M., Bugiel, S.: Short text, large effect: Measuring the impact of user reviews on android app security and privacy (2019).https: //doi.org/10.1109/SP.2019.00012

arXiv 2019
[19]

Olmstead, K., Atkinson, M.: Apps permissions report (2015),https://www.pewr esearch.org/wp-content/uploads/sites/9/2015/11/PI_2015-11-10_apps-per missions_FINAL.pdf

2015
[20]

OpenAI: Gpt-4 (2023),https://openai.com/product/gpt-4

2023
[21]

Palomba, F., Salza, P., Ciurumelea, A., Panichella, S., Gall, H., Ferrucci, F., Lucia, A.: Recommending and localizing change requests for mobile apps based on user reviews (2017).https://doi.org/10.1109/ICSE.2017.18

work page doi:10.1109/icse.2017.18 2017
[22]

Panichella, S., Di Sorbo, A., Guzman, E., Visaggio, C.A., Canfora, G., Gall, H.: How can i improve my app? classifying user reviews for software maintenance and evolution (2015).https://doi.org/10.1109/ICSM.2015.7332474

arXiv 2015
[23]

Pistoia, M., Tripp, O., Lubensky, D.: Combining static code analysis and machine learning for automatic detection of security vulnerabilities in mobile apps (2017)

2017
[24]

iScience 25(11), 105331 (2022).https://doi.org/10.1016/j.isci.2022.105331

Rajotte, J.F., Bergen, R., Buckeridge, D.L., El Emam, K., Ng, R., Strome, E.: Synthetic data as an enabler for machine learning applications in medicine. iScience 25(11), 105331 (2022).https://doi.org/10.1016/j.isci.2022.105331

arXiv 2022
[25]

ACM Comput

Sabir, B., Ullah, F., Babar, M.A., Gaire, R.: Machine learning for detecting data exfiltration: A review. ACM Comput. Surv.54(3) (May 2021).https://doi.org/ 10.1145/3442181,https://doi.org/10.1145/3442181

work page doi:10.1145/3442181 2021
[26]

Information and Software Technology122, 106290 (2020).https://doi.org/10.1016/j.infsof.2020.106290

Tao, C., Guo, H., Huang, Z.: Identifying security issues for mobile applications based on user review summarization. Information and Software Technology122, 106290 (2020).https://doi.org/10.1016/j.infsof.2020.106290

arXiv 2020
[27]

Expert Systems with Applications207, 117948 (2022).https://doi.org/https://doi.org/10.1016/j.eswa.2022.117948

Ullah, F., Ali Babar, M., Aleti, A.: Design and evaluation of adaptive system for big data cyber security analytics. Expert Systems with Applications207, 117948 (2022).https://doi.org/https://doi.org/10.1016/j.eswa.2022.117948

arXiv 2022
[28]

Journal of Systems and Software151, 81–118 (2019).https: //doi.org/https://doi.org/10.1016/j.jss.2019.01.051

Ullah, F., Ali Babar, M.: Architectural tactics for big data cybersecurity analytics systems: A review. Journal of Systems and Software151, 81–118 (2019).https: //doi.org/https://doi.org/10.1016/j.jss.2019.01.051

work page doi:10.1016/j.jss.2019.01.051 2019
[29]

Ullah, F., Edwards, M., Ramdhany, R., Chitchyan, R., Babar, M.A., Rashid, A.: Data exfiltration: A review of external attack vectors and countermeasures. Journal of Network and Computer Applications101, 18–54 (2018).https://doi.org/ht tps://doi.org/10.1016/j.jnca.2017.10.016,https://www.sciencedirect.co m/science/article/pii/S1084804517303569

work page doi:10.1016/j.jnca.2017.10.016 2018

[1] [1]

App store ratings and reviews (2019),https://developer.apple.com/app-store /ratings-and-reviews/

2019

[2] [2]

ACM Comput

Ahmad, H., Dharmadasa, I., Ullah, F., Babar, M.A.: A review on c3i systems’ security: Vulnerabilities, attacks, and countermeasures. ACM Comput. Surv.55(9) (Jan 2023).https://doi.org/10.1145/3558001,https://doi.org/10.1145/35 58001

work page doi:10.1145/3558001 2023

[3] [3]

AI and Society37, 1715–1728 (2021).https://doi.org/10.1007/s00146-021-0 1262-5

Andreotta, A.J., Kirkham, N., Rizzi, M.: Ai, big data, and the future of consent. AI and Society37, 1715–1728 (2021).https://doi.org/10.1007/s00146-021-0 1262-5

work page doi:10.1007/s00146-021-0 2021

[4] [4]

apple.com/en-gb/guide/iphone/iph251e92810/ios

Apple: Control access to information in apps on iphone (nd),https://support. apple.com/en-gb/guide/iphone/iph251e92810/ios

[5] [5]

arXiv preprint arXiv:1803.11175 (2018)

Cer, D., Yang, Y., Kong, S.y., Hua, N., Limtiaco, N., St John, R., Constant, N., Guajardo-Cespedes, M., Yuan, S., Tar, C., Sung, Y.H., Strope, B., Kurzweil, R.: Universal sentence encoder. arXiv preprint arXiv:1803.11175 (2018)

Pith/arXiv arXiv 2018

[6] [6]

International Journal of Innovative Research in Advanced Engineering3, 62–72 (2016).https://doi.org/ 10.6084/M9.FIGSHARE.3504380.V1

Chatterjee, S., Paul, K., Roy, R., Nath, A.: A comprehensive study on security issues in android mobile phone — scope and challenges. International Journal of Innovative Research in Advanced Engineering3, 62–72 (2016).https://doi.org/ 10.6084/M9.FIGSHARE.3504380.V1

work page doi:10.6084/m9.figshare.3504380.v1 2016

[7] [7]

In: Proceedings of ICSE

Chen, N., Lin, J., Hoi, S.C.H., Xiao, X., Zhang, B.: Arminer: mining informative reviews for developers from mobile app marketplace. In: Proceedings of ICSE. pp. 767–778 (2014)

2014

[8] [8]

Cho, J., Lee, K., Shin, E., Choy, G., Do, S.: How much data is needed to train a medical image deep learning system to achieve necessary high accuracy? arXiv preprint arXiv:1511.06348 (2016)

Pith/arXiv arXiv 2016

[9] [9]

arXiv preprint arXiv:1810.04805 (2019)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirec- tional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2019)

Pith/arXiv arXiv 2019

[10] [10]

Eason: Google play store apps reviews scraping and text analytics (sentiment anal- ysis) (2021),https://medium.com/analyticsvidhya/google-play-store-app s-reviews-scraping-and-text-analyticssentiment-analysis-5303294fffa7

2021

[11] [11]

edu/soups/2012/proceedings/a3_Felt.pdf

Felt, A., Ha, E., Egelman, S., Haney, A., Chin, E., Wagner, D.: Android permis- sions: User attention, comprehension, and behavior (nd),https://cups.cs.cmu. edu/soups/2012/proceedings/a3_Felt.pdf

2012

[12] [12]

Grano, G., Di Sorbo, A., Mercaldo, F., Visaggio, C.A., Canfora, G., Panichella, S., Visaggio, C.: Android apps and user feedback: A dataset for software evolution and quality improvement (2017).https://doi.org/10.1145/3121264.3121266

arXiv 2017

[13] [13]

ACM Computing Surveys55(1), 1–36 (2021).https://doi.org/10.1145/3487890

Hu, Y., Kuang, W., Qin, Z., Li, K., Zhang, J., Gao, Y., Li, W., Li, K.: Artificial in- telligence security: Threats and countermeasures. ACM Computing Surveys55(1), 1–36 (2021).https://doi.org/10.1145/3487890

work page doi:10.1145/3487890 2021

[14] [14]

arXiv (2022)

Li, Y., Dang, X., Tian, H., Sun, T., Wang, Z., Ma, L., Klein, J., Bissyande, T.: Ai-driven mobile apps: an explorative study. arXiv (2022)

2022

[15] [15]

arXiv (2021)

Lu, Y., Wang, W., Wei, W., Wang, H.: Machine learning for synthetic data gener- ation: a review. arXiv (2021)

2021

[16] [16]

Mukherjee, D., Ahmadi, A., Pour, M., Reardon, J.: An empirical study on user reviews targeting mobile apps’ security privacy (2020)

2020

[17] [17]

In: Proceedings of the 44th International Conference on Software Engineering (2022).https://doi.org/10.1145/3510003.3510079

Nema, P., Anthonysamy, P., Taft, N., Peddinti, S.T.: Analyzing user perspectives on mobile app privacy at scale. In: Proceedings of the 44th International Conference on Software Engineering (2022).https://doi.org/10.1145/3510003.3510079

arXiv 2022

[18] [18]

Nguyen, D.C., Derr, E., Backes, M., Bugiel, S.: Short text, large effect: Measuring the impact of user reviews on android app security and privacy (2019).https: //doi.org/10.1109/SP.2019.00012

arXiv 2019

[19] [19]

Olmstead, K., Atkinson, M.: Apps permissions report (2015),https://www.pewr esearch.org/wp-content/uploads/sites/9/2015/11/PI_2015-11-10_apps-per missions_FINAL.pdf

2015

[20] [20]

OpenAI: Gpt-4 (2023),https://openai.com/product/gpt-4

2023

[21] [21]

Palomba, F., Salza, P., Ciurumelea, A., Panichella, S., Gall, H., Ferrucci, F., Lucia, A.: Recommending and localizing change requests for mobile apps based on user reviews (2017).https://doi.org/10.1109/ICSE.2017.18

work page doi:10.1109/icse.2017.18 2017

[22] [22]

Panichella, S., Di Sorbo, A., Guzman, E., Visaggio, C.A., Canfora, G., Gall, H.: How can i improve my app? classifying user reviews for software maintenance and evolution (2015).https://doi.org/10.1109/ICSM.2015.7332474

arXiv 2015

[23] [23]

Pistoia, M., Tripp, O., Lubensky, D.: Combining static code analysis and machine learning for automatic detection of security vulnerabilities in mobile apps (2017)

2017

[24] [24]

iScience 25(11), 105331 (2022).https://doi.org/10.1016/j.isci.2022.105331

Rajotte, J.F., Bergen, R., Buckeridge, D.L., El Emam, K., Ng, R., Strome, E.: Synthetic data as an enabler for machine learning applications in medicine. iScience 25(11), 105331 (2022).https://doi.org/10.1016/j.isci.2022.105331

arXiv 2022

[25] [25]

ACM Comput

Sabir, B., Ullah, F., Babar, M.A., Gaire, R.: Machine learning for detecting data exfiltration: A review. ACM Comput. Surv.54(3) (May 2021).https://doi.org/ 10.1145/3442181,https://doi.org/10.1145/3442181

work page doi:10.1145/3442181 2021

[26] [26]

Information and Software Technology122, 106290 (2020).https://doi.org/10.1016/j.infsof.2020.106290

Tao, C., Guo, H., Huang, Z.: Identifying security issues for mobile applications based on user review summarization. Information and Software Technology122, 106290 (2020).https://doi.org/10.1016/j.infsof.2020.106290

arXiv 2020

[27] [27]

Expert Systems with Applications207, 117948 (2022).https://doi.org/https://doi.org/10.1016/j.eswa.2022.117948

Ullah, F., Ali Babar, M., Aleti, A.: Design and evaluation of adaptive system for big data cyber security analytics. Expert Systems with Applications207, 117948 (2022).https://doi.org/https://doi.org/10.1016/j.eswa.2022.117948

arXiv 2022

[28] [28]

Journal of Systems and Software151, 81–118 (2019).https: //doi.org/https://doi.org/10.1016/j.jss.2019.01.051

Ullah, F., Ali Babar, M.: Architectural tactics for big data cybersecurity analytics systems: A review. Journal of Systems and Software151, 81–118 (2019).https: //doi.org/https://doi.org/10.1016/j.jss.2019.01.051

work page doi:10.1016/j.jss.2019.01.051 2019

[29] [29]

Ullah, F., Edwards, M., Ramdhany, R., Chitchyan, R., Babar, M.A., Rashid, A.: Data exfiltration: A review of external attack vectors and countermeasures. Journal of Network and Computer Applications101, 18–54 (2018).https://doi.org/ht tps://doi.org/10.1016/j.jnca.2017.10.016,https://www.sciencedirect.co m/science/article/pii/S1084804517303569

work page doi:10.1016/j.jnca.2017.10.016 2018