PaperlayerView exact revision on arXiv

Source-linked AI summary

The Feasibility of Dynamically Granted Permissions: Aligning Mobile Privacy with User Preferences

Primal Wijesekera, Arjun Baokar, Lynn Tsai, Joel Reardon, Serge Egelman, David Wagner, Konstantin Beznosov

arXiv:1703.02090v1cs.CR

TL;DR

Ask-on-first-use permissions overlook the context of recurring requests, weakening alignment with user preferences. This paper studies contextual privacy decisions and builds a classifier using context and behavior, reporting 96.8% accuracy in predicting users’ decisions.

  • Problem

    Existing ask-on-first-use permission systems ignore contextual information that shapes users’ privacy decisions.

  • Method

    The study builds an ML classifier using permission decisions, user behavior, and request context to infer privacy preferences.

  • Results

    The classifier accounted for contextual information and passively observable behavior, supporting both corresponding research hypotheses.

  • Takeaways & Limitations

    Permission-granting systems have significant room for improvement and may benefit from incorporating contextual and behavioral information.

  • Takeaways & Limitations

    The study shows that ask-on-first-use still violates user privacy around one in seven permission requests when users deviate from their initial responses.

Abstract

from arXiv · show

Current smartphone operating systems regulate application permissions by prompting users on an ask-on-first-use basis. Prior research has shown that this method is ineffective because it fails to account for context: the circumstances under which an application first requests access to data may be vastly different than the circumstances under which it subsequently requests access. We performed a longitudinal 131-person field study to analyze the contextuality behind user privacy decisions to regulate access to sensitive resources. We built a classifier to make privacy decisions on the user's behalf by detecting when context has changed and, when necessary, inferring privacy preferences based on the user's past decisions and behavior. Our goal is to automatically grant appropriate resource requests without further user intervention, deny inappropriate requests, and only prompt the user when the system is uncertain of the user's preferences. We show that our approach can accurately predict users' privacy decisions 96.8% of the time, which is a four-fold reduction in error rate compared to current systems.

I. INTRODUCTION

Existing mobile permission systems prompt users without adequately accounting for the context surrounding data access, risking decisions misaligned with user preferences. This study investigates context-aware, machine-learned predictions to reduce unnecessary prompts while preserving user control.

  • Motivation: Ask-on-first-use improves on ask-on-install by providing runtime context, but still may not reflect later circumstances surrounding an application’s data requests.Earlier install-time models obscured context and were poorly understood; runtime prompting gives users more information about what they were doing when access was first requested.
  • Motivation: Contextual integrity implies that privacy violations occur when sensitive resources are used in ways that defy users’ expectations.The system must balance prompting often enough to respect preferences without causing habituation or acting against users’ wishes.
  • Study design: 131 participants used instrumented Android phones for an average of 32 days to provide real-world data for inferring future privacy decisions.The study examined past decisions, application-request contexts, and behavioral traits, with phones periodically prompting participants about permission requests.
  • Approach: The classifier uses related past decisions to predict responses, returning a confidence score and prompting users only when confidence is insufficient.This design aims to avoid prompting in most cases and reduce user burden while retaining user input for uncertain decisions.
  • Contributions: A significant portion of participants made contextual permission decisions based on the foreground application and the requesting application’s visibility.The work also presents machine-learning techniques that incorporate context and passively observed traits to infer future decisions case by case at runtime.

II. RELATED WORK

Prior work shows that conventional permission prompts often fail to support informed, context-sensitive privacy decisions. Researchers have explored finer-grained controls, privacy-preference inference, contextual integrity, and machine learning, but existing approaches can increase user involvement, rely on limited datasets, or use static recommendations.

  • Limitations of Existing Permission Models: Install-time prompts often fail because users misunderstand permission types, overlook them, and are surprised by background data collection.Prior work also reports that applications transmit substantial amounts of location and other sensitive data to third parties.
  • Limitations of Existing Permission Models: Finer-grained permission and information-flow systems can increase user involvement, cause habituation, and mainly benefit motivated or technically savvy users.Some proposals require users to configure complicated control panels.
  • Privacy-Preference Inference: Liu et al. used machine learning to predict whether users would allow or deny sensitive-data requests, but studied highly privacy-conscious users and removed conflicting decisions.The dataset came from individuals who chose to install a permission-control mechanism.
  • Contextual Privacy Models: A Privacy Assistant increased awareness through privacy-profile recommendations, but its recommendations were static and did not account for each application’s sensitive-data access on a case-by-case basis.The assistant combined inferred privacy profiles with static analysis of third-party applications.
  • Contextual Privacy Models: Contextual-integrity research identifies request context, requesting-agent role, and resource type as components of appropriate information flows, while field studies found application visibility strongly influenced decisions and recommended machine learning for individual preferences.Observed privacy decisions were highly nuanced, making a one-size-fits-all model unlikely to suffice.

III. METHODOLOGY · A. Instrumentation

The study combined smartphone instrumentation with in situ Experience Sampling Method probes to collect behavioral, runtime, and user-decision data about privacy preferences. Across a six-week deployment with 131 participants, the instrumentation recorded permission use and request context while sampling frequent requests for user feedback.

  • III. METHODOLOGY: 131 participants provided data for inferring which factors predict whether users consider permission requests appropriate.
  • III. METHODOLOGY: The study focused on 10 permission types after excluding permissions considered better handled by other mechanisms.
  • III. METHODOLOGY: Experience Sampling Method probes asked participants in situ whether they would have permitted an application’s recent data access, treating responses as the main dependent variable.
  • III. METHODOLOGY: The Android instrumentation required operating-system modifications and flashing a custom Android version onto participants’ smartphones.
  • III. METHODOLOGY: The six-week over-the-air deployment ran between February 2016 and March 2016, followed by an exit survey and institutional review board approval.
  • A. Instrumentation: Instrumentation collected behavioral information, runtime information, and user decisions with minimal performance impact and without modifying third-party application code.
  • A. Instrumentation: Behavioral features covered web browsing, screen locking, application use, audio, calls, camera usage, and security-setting behaviors.
  • A. Instrumentation: Runtime records captured each permission request’s application visibility and the foreground application representing the user’s activity.

B. Exit Survey · C. Summary · IV. TYPES OF USERS

The study combined participant survey data with a large-scale permission-event dataset, then distinguished users by denial rate and sensitivity to application visibility. Contextuals incorporated visibility into privacy decisions, whereas Defaulters generally did not.

  • B. Exit Survey: The exit survey collected demographic information and qualitative information about participants’ privacy preferences.Participants were incentivized with a raffle for two $100 Amazon gift cards.
  • B. Exit Survey: Of 203 experiment participants, 53 fully completed the exit survey and 14 partially completed it.Among full respondents, 21 were male, 31 female, and 1 undisclosed; ages ranged from 20 to 72 years (µ = 40.83, σ = 14.32).
  • C. Summary: 131 participants used the custom Android build for more than 20 days, generating 176M events across 1,686 applications and 13K activities.Participants responded to 4,636 prompts during the study period.
  • C. Summary: 55.3% of 4,224 prompts involved ACCESS _WIFI_STATE, while participants wanted to deny 60.01% of prompts.Other prompt categories included direct location access (21.0%), reading SMS (17.3%), sending SMS (5.08%), reading call logs (0.78%), and browser history (0.54%).
  • IV. TYPES OF USERS: The study grouped users by denial rate in 10% increments because its sample was too small for effective clustering.The researchers used denial rate as a meaningful distinction between user sub-populations.
  • IV. TYPES OF USERS: Application visibility significantly affected users with denial rates of 10–90%, but not users with denial rates of 0–10% or 90–100%.The researchers labeled the former group Contextuals and the latter group Defaulters.
  • IV. TYPES OF USERS: 53% of 131 participants were Defaulters and 47% were Contextuals.Contextuals showed a significant response difference by application visibility (p < 0.013, r = 0.312), whereas Defaulters did not (p = 0.227, r = 0.215).

V. ASK-ON-FIRST-USE PERMISSIONS · AOFU-AP · AOFU-AF PV

The section evaluates ask-on-first-use permission policies, comparing application–permission prompting with richer contextual combinations and examining privacy violations, functionality losses, and differences between user types. It also notes that deployed AOFU may overestimate effectiveness because it prompts for requests that do not actually expose sensitive data.

  • V. ASK-ON-FIRST-USE PERMISSIONS: AOFU prompts users on the first dangerous-permission request for each application–permission pair, then reuses that response for later requests.AOFU was introduced in Android 6.0, but only 24.3% of Android users had Marshmallow or higher as of November 2016.
  • AOFU-AF PV: Table III reports expected median accuracy for alternative ask-on-first-use combinations, including policies that incorporate application visibility and foreground application context.The reported values are medians except average prompt count, because the distributions were skewed.
  • AOFU-AP: The analysis uses AOFU-AP as the Android Marshmallow baseline and considers whether prompting on application–permission–visibility combinations changes accuracy and prompt frequency.The policy dimensions include the requesting application, permission type, application visibility, and foreground application.
  • V. ASK-ON-FIRST-USE PERMISSIONS: 24.3% of Android users had Marshmallow or higher by November 2016, while the remaining 75.7% used ask-on-install, which automatically allows runtime permission requests.All study participants had ask-on-install as their default permission model.
  • AOFU-AF PV: A permission allowed against the user’s preference is a privacy violation, whereas a permission denied despite the user’s preference causes functionality loss.Under ask-on-install, every denied ESM prompt corresponds to a misprediction because the model grants access regardless of the user’s preference.
  • AOFU-AF PV: AOFU performs better for Defaulters than Contextuals because Defaulters remain consistent after initial responses, while Contextuals vary with factors beyond application and permission type.This difference shows that application–permission history alone does not fully capture contextual privacy preferences.
  • AOFU-AF PV: Deployed Marshmallow AOFU may overestimate effectiveness because it prompts for permission requests that check authorization without actually revealing sensitive data.The study’s simulation avoids such false positives by prompting only when the invoked function accesses sensitive data.
  • AOFU-AF PV: AOFU-AF PV can have higher median accuracy for Defaulters and Contextuals separately, but its overall median accuracy equals AOFU-AP when the groups are combined.The equality is attributed to skewed distributions.

VI. LEARNING PRIVACY PREFERENCES · A. Feature Selection

The study models privacy decisions using behavioral, contextual, runtime, and aggregate features because 47% of users make decisions beyond application, permission, and visibility. Feature selection reduced 16K candidate combinations to 20 features, including prior denial-rate aggregates that capture user-specific preferences.

  • VI. LEARNING PRIVACY PREFERENCES: 47% of users were classified as Contextuals, making privacy decisions based on factors beyond the requesting application, permission, and application visibility.The model was trained and tested on 4,224 prompts collected from 131 users.
  • VI. LEARNING PRIVACY PREFERENCES: The approach assumes a trusted Android platform enforces permissions and applications cannot bypass its permission system.It also assumes a non-adversarial machine-learning setting in which attackers do not exploit the classifier’s decision process.
  • A. Feature Selection: 16K candidate features were generated from behavioral, contextual, and aggregate combinations of specific applications and actions, then reduced to 20 features.Selection used random-forest Gini importance, correlation significance testing, and singular value decomposition to address sparsity and high dimensionality.
  • A. Feature Selection: Behavioral features predictive of privacy decisions covered browsing habits, audio and call traits, and locking behavior.Examples included websites visited, HTTPS-link proportion, downloads, location-requesting sites, and passcode-locking behavior; values were normalized per day/user and scaled.
  • A. Feature Selection: Runtime features included requesting-application visibility, requested permission, and request time of day, while application identity and popularity were not predictive.User ID was initially included to capture user-to-user variance but was subsequently removed.
  • A. Feature Selection: The model needed to determine how each user treats permission requests because users may perceive the same permission threat differently.This motivates features that capture user-specific privacy preferences for predicting future decisions.
  • A. Feature Selection: Aggregate features captured prior responses by computing average denial rates for application:permission:visibility and permission:applicationF 4:visibility combinations.These runtime-measurable aggregates were designed to partly capture differences in users’ privacy preferences.

B. Inference Based on Behavior · C. Inference Based on Contextual Cues

The paper evaluates privacy-decision inference from passively observed behavior and contextual cues, including models requiring no new user input. The best contextual model reduced errors substantially, but required more prompts than Android’s existing approach.

  • B. Inference Based on Behavior: Behavioral inference used passively collected behavioral and runtime data to predict future privacy decisions with an SVM model.The behavioral categories included privacy and security settings, locking, audio and call habits, browsing, and application usage.
  • B. Inference Based on Behavior: 83% cross-validation accuracy was achieved across users using behavioral and runtime features, assuming prior privacy decisions from each user.This setup was primarily relevant after users had used their phones for some time.
  • B. Inference Based on Behavior: 75% median accuracy was achieved in leave-one-out testing, predicting each user from the other 130 participants’ behavioral and runtime data.This setup required no prompts or other effort from the held-out user and improved threefold over AOI.
  • C. Inference Based on Contextual Cues: Contextual features improved predictions for users who varied decisions, while runtime features alone were highly accurate for Defaulters who followed default allow-or-deny policies.Aggregate features captured consistency across application, permission, and visibility combinations.
  • C. Inference Based on Contextual Cues: 15.38% to 7.76% was the reduction in error rate achieved by the selected R2, A1, and A2 model versus AOFU-AP.The selected model also had the best accuracy and fewest privacy violations among the evaluated feature sets.
  • C. Inference Based on Contextual Cues: 24 prompts per user were required on average by the model, compared with 12 for AOFU-AP, because the model could prompt repeatedly for the same combination.Increasing prompts improved accuracy until a steady state, after which additional prompts plateaued; some users remained difficult to infer.

VII. LEARNING STRATEGY · A. Bootstrapping

The learning strategy aims to infer a new user’s privacy preferences with minimal prompting, then use model confidence to balance user involvement and accuracy. During bootstrapping, the model leverages other users’ data and progressively incorporates the new user’s prompt responses, improving accuracy while reducing prompts relative to existing policies.

  • VII. LEARNING STRATEGY: The learning phase seeks to learn most privacy preferences with minimal user involvement.After adequate training, model confidence can assess performance across users and the tradeoff between involvement and accuracy.
  • VII. LEARNING STRATEGY: Class probabilities were calculated using the KSVM library in R with Platt et al.’s probability-estimation technique.
  • A. Bootstrapping: During bootstrapping, a model trained on other users’ preferences predicts privacy decisions for a new user with no prior model information.The model is evaluated as prompts are progressively presented to that user.
  • A. Bootstrapping: 66.6% median accuracy was achieved in leave-one-out validation, using prompt responses from 130 users to predict the remaining user.Accuracy was 56.2% for Contextuals and 86.4% for Defaulters.
  • A. Bootstrapping: 96.6% accuracy was achieved for Defaulters before six prompts, exceeding AOFU’s 93.33% accuracy from the beginning.This reduced permission prompts by 83.3% compared to AOFU-AP.
  • A. Bootstrapping: 80% error reduction over AOFU-AP was achieved after 12 user prompts, while the strategy required an average of nine prompts to reach maximum capacity.That average represented a 25% reduction in user involvement over AOFU-AP.
  • A. Bootstrapping: Contextuals required more prompts because their decisions vary with contextual cues, whereas Defaulters are easier to learn from consistent early decisions.The authors therefore emphasize differentiating these groups early in bootstrapping.
  • A. Bootstrapping: 7.6% versus 3.2% error rates were obtained with learning-strategy selection versus random training-set selection, while prompts fell from 24 to 9.The hybrid AOFU-style bootstrapping approach achieved higher accuracy than AOFU with significantly fewer prompts.

B. Decision Confidence

After 12 AOFU-style prompts, the model assigned high-confidence predictions for nearly all users and substantially reduced errors with fewer prompts. Confidence also identified users for whom the approach may perform poorly, enabling prompt limits or fallback to AOFU.

  • Limitations: For a small subset of users, additional permission prompts did not improve accuracy, possibly because decisions were random or the model’s features were not predictive.The authors therefore treated confidence as a way to identify users for whom the system may not perform well.
  • Confidence measurement: Predictions with probability >0.60 were labeled high-confidence, while those with probability <0.60 were labeled low-confidence.Confidence was based on the binary classifier’s decision probability, with values closer to 0.5 indicating lower confidence.
  • Confidence measurement: 10 users (7.63% of 131 users) had at least one low-confidence prediction after 12 AOFU-style prompts, while 92.37% had only high-confidence predictions.The model was trained on other users’ responses and each user’s first 12 AOFU-selected prompts.
  • Accuracy and prompting: Among users with no low-confidence predictions, the hybrid approach achieved a 2% median error rate versus 13.3% for AOFU after 12 prompts.AOFU required an average of 15.11 prompts to reach its accuracy, whereas the ML model required 9 prompts on average.
  • Accuracy and prompting: The ML model reduced error rate by 84.61% compared with AOFU despite using fewer prompts in the learning phase.The approach worked well for the majority of users, while users with poor performance could be given limited prompts or use AOFU instead.

C. Online Model · VIII. CONTEXTUAL INTEGRITY

The online model incrementally adapts permission classifiers using phone-derived features and trusted-server training. The contextual-integrity analysis explains permission failures as omissions of application visibility and request context, while presenting dynamic regulation as an initial step that needs further study.

  • C. Online Model: The system trains models on a trusted server, distributes weight vectors to phones, and uses online learning to adapt to changing user preferences.Training across multiple users also provides more labeled data points.
  • C. Online Model: The implementation combines on-phone feature extraction and storage with a trusted-server machine-learning pipeline that exchanges feature vectors and newly trained classifier weights.The server can bootstrap models using a model trained on a few hundred users.
  • C. Online Model: The online classifier was evaluated with stochastic gradient descent and five-fold cross validation on a 4,224-point data set, followed by timestamp-ordered simulation of incoming data.Time-varying features were recomputed for each incoming data point to create phone-side feature snapshots.
  • C. Online Model: The trusted server receives no personally identifiable data, and transmitted features are scaled as standard deviations from the mean rather than reported as raw values.These measures address the concern that a trusted server could be compromised.
  • VIII. CONTEXTUAL INTEGRITY: Contextual integrity explains permission-model failures by emphasizing the context surrounding users’ privacy decisions, which the proposed ML model attempts to detect when it changes.The model aims to avoid prompting when a user’s sharing preference remains appropriate under the same circumstances.
  • VIII. CONTEXTUAL INTEGRITY: In the contextual-integrity framework, agents, roles, contexts, knowledge states, and communication actions describe whether transferring requested sensitive information is acceptable.Android permissions can be viewed as transferring knowledge of a guarded resource from the user/platform to a third-party application.
  • VIII. CONTEXTUAL INTEGRITY: Application visibility helps users infer the requesting application’s role, while the foreground application establishes context and their combination influences whether resource sharing is expected.A visible Google Maps navigation request may be acceptable whereas an invisible background service request may not be.
  • VIII. CONTEXTUAL INTEGRITY: Because AOFU controls access mainly through application-permission combinations, it neglects application visibility and context and is wrong in nearly one-fifth of cases.The authors describe further research on defining context as necessary, while characterizing dynamic permission regulation as an initial step.

IX. DISCUSSION · A. Limitations of Permission Models · B. Our ML-Based Model

The discussion finds that deployed Android permission models often misalign with users’ privacy preferences because they ignore contextual variation. An ML-based model using prior decisions, behavior, and environmental context substantially improves permission prediction while reducing user burden.

  • IX. DISCUSSION: The study tested four hypotheses concerning AOFU privacy violations, the usefulness of contextual information, ML-based improvement, and behavioral predictors.The authors report that hypotheses (i) and (ii) held in the field study, while the classifier results supported (iii) and (iv).
  • A. Limitations of Permission Models: The field study found that 95% of participants blocked at least one permission request, with an average denial rate of 60%.The discussion interprets this denial rate as evidence that users care about privacy and are wary of requests violating their expectations.
  • A. Limitations of Permission Models: AOI correctly regulated permission requests only one in four times because it permitted all requests after application installation.This contrasts the observed denial behavior with Android’s earlier install-time permission model.
  • A. Limitations of Permission Models: AOFU still violated user privacy around one in seven times when users deviated from their initial permission responses.The discussion attributes this limitation to AOFU’s assumption that decisions depend only on the application:permission combination.
  • B. Our ML-Based Model: The ML techniques learned from previous permission decisions and current environmental context to predict grants on users’ behalf with less burden.The authors report better results than methods deployed on millions of phones worldwide.
  • B. Our ML-Based Model: 75%: the ML model reduced AOFU errors; its 97% accuracy exceeded AOFU’s 85% and AOI’s 25%.These methods comprise the status quo in the Android ecosystem.
  • B. Our ML-Based Model: Environmental context significantly influenced users’ permission decisions, which were often neither random nor fixed.The authors argue that systems need not infer the precise context of data use to improve privacy support, because environmental data can still improve decisions.

C. Reducing the User Burden … F. Experimental Caveat

The proposed model reduces user burden while maintaining or improving privacy-decision accuracy, including through passive behavioral inference. The remaining sections examine tailored models, attack risks, and experimental limitations involving the consequences of denied permissions.

  • C. Reducing the User Burden: Three-fold improvement over AOI was achieved by a behavioral-traits-only model, which matched AOFU-AP accuracy for Defaulters comprising 53% of the sample.The approach inferred privacy preferences from passively observable behavior without permission prompts.
  • C. Reducing the User Burden: 81% reduction in prediction errors and 25% less user involvement were achieved by combining contextual data with users’ past decisions.This improvement over AOFU may reduce habituation and improve reliability when user input is needed.
  • D. User- and Permission-Tailored Models: User- and permission-tailored models were evaluated as a possible alternative to the single model trained on data from all users.Users were clustered first by behavioral features and then by denial rate to test whether specialized models could outperform the general-purpose model.
  • D. User- and Permission-Tailored Models: Tailored models may become more useful when the permission system is implemented at scale.The passage presents this as a possibility rather than an established result.
  • E. Attacking the ML Model: Attacks could target the permission model through adversarial environments, training-set pollution, or manipulated input features.The analysis assumes that the platform is not compromised; a compromised platform would degrade permission systems generally.
  • E. Attacking the ML Model: Preliminary analysis found that adversarial access scenarios may be non-trivial, but comprehensive study and defenses against adversarial ML and feature brittleness remain necessary.Behavioral features cannot be influenced without compromising the platform, whereas runtime features can be controlled by choosing when to request permission.
  • F. Experimental Caveat: Users in the experiment could deny permissions without consequences, so denial rates measured privacy preferences and expectations rather than tradeoffs against application functionality.Actual denial could lower denial rates, and the study leaves replication under real permission consequences for future work.
  • F. Experimental Caveat: Future platforms should provide applications with apparently allowed permissions while supplying only spurious or artificial data to avoid forcing choices between privacy and functionality.This direction addresses usability concerns because many applications may crash when permissions are denied.

G. Types of Users · H. User Interface Panel · I. Conclusions

The paper categorizes users by how application visibility shapes privacy decisions, proposes a feedback panel to review and correct model decisions, and concludes that permission choices depend on context beyond application and permission identity. It reports 97% accuracy while acknowledging prediction risks and opportunities for improvement.

  • G. Types of Users: Users are categorized by the significance of application visibility in their individual privacy decisions.The observed category distribution may differ in an actual permission-denial setting.
  • G. Types of Users: Visibility is proposed as an important factor distinguishing Contextuals from Defaulters.The categorization motivates deeper analysis of the factors separating these groups.
  • H. User Interface Panel: 97% accuracy remains improvable because any model predicting user decisions can make incorrect predictions.Prediction is necessary because applications request permissions too frequently for users to examine manually.
  • H. User Interface Panel: A user feedback panel could let users review decisions made by the ML model and adjust them to their preferences.This provides recourse for correcting errors and undesirable decisions.
  • I. Conclusions: Users care about privacy and deny a significant number of requests to access sensitive data.This is one of the paper’s reported conclusions.
  • I. Conclusions: Existing Android permission models still result in significant privacy violations.The conclusion identifies shortcomings in current permission handling.
  • I. Conclusions: Users may allow permissions at some times and deny them at others, showing that decisions involve more than application name and permission type.The finding indicates that privacy decisions vary across circumstances.
  • I. Conclusions: 131 users provided real-world data showing that application visibility and the current foreground application were important factors in user decisions.These factors were identified from the collected field data.

APPENDIX A INFORMATION GAIN OF FEATURES · APPENDIX B INFORMATION GAIN OF BEHAVIORAL FEATURES

The supplied passages identify appendices on feature and behavioral-feature information gain, but provide no accompanying results or feature-level values.

  • APPENDIX B INFORMATION GAIN OF BEHAVIORAL FEATURES: Appendix B includes a table titled “FEATURE IMPORTANCE OF BEHAVIORAL FEATURES.”The supplied table excerpt provides the title only, without feature rankings or information-gain values.
Loading 1703.02090v1…