Mobile App Privacy Detection via ML Feature Vectors
Solution Overview
Problem
Existing mobile platforms lack effective mechanisms to distinguish between legitimate and illegitimate usage of private user data, leading to privacy violations, as current techniques like information-flow tracking and data-centric analysis are brittle and prone to false positives and cannot handle custom data transformations.
Innovation Solution
A method that mines regular histories of mobile application programs to extract learnable features, combines them into a single feature vector, and uses a machine-learning-based classification algorithm to determine unauthorized privacy data access, incorporating call-graph representations and abstracting sequences of events to improve privacy violation detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If information-flow tracking (taint analysis) is used to monitor privacy data flow, then privacy leakage detection capability is improved, but the system becomes brittle and produces false positives when encountering covert channels, implicit flows, or custom data transformations
Solution Approach 1:
The patent replaces the mechanical taint analysis approach with a machine learning classification system. Instead of manually tracking data flow through code analysis, the system automatically learns privacy violation patterns from training data, substituting rigid mechanical analysis with adaptive statistical learning that handles custom transformations and covert channels more robustly
Solution Approach 2:
The patent transforms the detection approach by changing parameters from binary taint tags to continuous feature vectors containing multiple attributes (data flow paths, transformation types, destination categories). This parameter transformation enables the machine learning classifier to capture nuanced patterns that simple reachability queries miss, reducing false positives while maintaining detection accuracy
2Device complexity
If data-centric analysis is used to compare source and sink values, then instrumentation overhead is reduced, but false alarms increase due to benign values and inability to handle custom data transformations
Solution Approach 1:
The patent creates a universal machine learning classification framework that handles multiple detection scenarios simultaneously - it can detect both direct data flows and transformed data, handle various transformation types (encryption, hashing, custom functions), and work across different privacy sources and sinks. This multi-functional approach eliminates the need for separate analysis methods for different cases, reducing overall system complexity while improving accuracy
Solution Approach 2:
The patent performs preliminary action by training the machine learning classifier on extensive training data that includes both benign and malicious applications with various data transformation patterns. This pre-learning phase enables the system to automatically distinguish benign from malicious values without requiring complex runtime analysis, reducing instrumentation overhead while maintaining high detection accuracy
3Ease of operation
If permission model is used to control access to private information, then basic access control is provided, but the system cannot disambiguate legitimate from illegitimate usage once permission is granted
Solution Approach 1:
The patent introduces a machine learning classification system as an intermediary layer between the permission model and actual data access. The classifier analyzes application behavior patterns and predicts whether access is legitimate or illegitimate, serving as a mediator that enhances the basic permission model without requiring complete redesign of the access control mechanism
Solution Approach 2:
The patent implements feedback by using the machine learning classifier to continuously evaluate whether permission usage aligns with legitimate patterns. The system learns from training data what constitutes appropriate usage and provides feedback signals that can trigger alerts or enforcement actions when illegitimate patterns are detected, enabling dynamic adjustment of access control decisions
Data Source
AI summary
Privacy violation detection of a mobile application program is disclosed. Regular histories of the mobile application are mined. A call-graph representation of the mobile application program can be created and sequences of events of interest according to the platform specification of the mobile application can be collected. A plurality of learnable features are extracted from the regular histories. The plurality of learnable features are combined into a single feature vector which is fed into a machine-learning-based classification algorithm. Whether the mobile application program includes one or more permissions for accessing unauthorized privacy data of a mobile application user is determined based on a machine learning classification of the single feature vector. The collected sequences can be reduced into a plurality of feature vectors which can include at least one of a happens-before feature and a multiplicity of occurrences feature.