Security Evaluation Research of Mobile Payment Application Based on SVM

Yang Liu, Juan Xu, Hui Fei, Yanhui Guo, Guoai Xu2016 · DOI

Abstract

With the increase of mobile payment malicious applications, an accurate assessment of their safety is particularly important. In recent years, machine learning methods have achieved good results in text recognition, medical diagnosis, anomaly detection and other fields. Therefore, we consider the introduction of the application of many features including application signature information, application permissions, suspicious API, special string to construct a sample feature space and evaluate the safety of mobile payment applications based on support vector machine algorithm. By comparing the accuracy of a variety of algorithms, the algorithm is turned to be feasible. The importance of the evaluation index is different during the sample classification and there is a difference in the performance of the kernel function under different conditions. So this paper studies the performance of the classifier in different evaluation index set, kernel function and feature weights. It proves that the polynomial kernel support vector machine method is best based on feature weights after the introduction of four types of evaluation index. This method properly avoids the problem of low accuracy due to a single feature while eliminating the adverse effects of weak correlation evaluation index.

Introduction

As mobile payment becomes ubiquitous, it has become a prime target for "secondary packaging" attacks. In these scenarios, hackers take a legitimate banking or payment app, inject malicious code (viruses or trojans), and re-release it on third-party markets. Traditional security research often focuses heavily on data encryption and authentication protocols while ignoring the security of the application itself. Because these malicious apps often look and act like the originals, they are difficult for users to spot. This research addresses the challenge of creating an automated system that can analyze the "DNA" of an app—its requested permissions and internal code behavior—to determine if it has been tampered with.

Methodology

The proposed multi-feature evaluation model utilizes Support Vector Machines (SVM) to distinguish between normal and malicious applications. The process follows a systematic pipeline:

  1. Feature Extraction: Using the open-source tool Androguard, four categories of features are extracted:
  2. S1 (Signature Information): Checks for alterations in the developer certificate.
  3. S2 (Permissions): Analyzes 13 specific Android privileges (e.g., SEND_SMS, READ_CONTACTS) frequently abused by malware.
  4. S3 (Suspicious APIs): Monitors calls to sensitive functions like installPackage() or Runtime.exec().
  5. S4 (Special Strings): Searches for hardcoded IP addresses, emails, or shell commands.
  6. Model Construction: These features form a sample feature space used to train the SVM classifier.
  7. Optimization: The model is refined by experimenting with different kernel functions (Linear, Polynomial, and Radial Basis Function) and applying feature weights to emphasize critical security indicators.
Overview of the security evaluation pipeline from feature extraction to machine learning classification.

Experimental Setup

The study utilized a dataset of 1,999 Android applications, consisting of 1,612 normal samples collected from reputable stores like Google Play and 387 malicious samples sourced from forums and malware repositories. Statistical analysis revealed a clear divide: malicious apps requested significantly more permissions (averaging 11.64 vs. 7.21) and made nearly double the suspicious API calls compared to legitimate apps.

Key Results and Discussion

The SVM model demonstrated superior performance compared to other common data mining methods. As shown in the results, the SVM achieved a baseline accuracy of 96.50% with a remarkably low deviation of 0.32%, indicating high stability.

Comparison of accuracy rates and deviations across different data mining methods.

Further optimization revealed several key insights:

  • Kernel Selection: While the Radial Basis Function (RBF) is often the default, the Polynomial kernel proved superior for mobile security (96.99% accuracy). This is attributed to its ability to account for global characteristics and interactions between distant features in the space.
  • Feature Weighting: By assigning higher weights to signature information and special strings, the final model reached a peak accuracy of 97.82%.
  • Synergy: The integration of all four feature sets (S1-S4) is essential, as no single feature serves as a "silver bullet" for detection.

Conclusion

This research proves the feasibility of using a weighted SVM model with a polynomial kernel for mobile payment security evaluation. While highly accurate, future work must address data imbalance through sample weighting and adapt to evolving malware obfuscation techniques that may challenge static feature extraction.

Summary updated

Abstract

With the increase of mobile payment malicious applications, an accurate assessment of their safety is particularly important. In recent years, machine learning methods have achieved good results in text recognition, medical diagnosis, anomaly detection and other fields. Therefore, we consider the introduction of the application of many features including application signature information, application permissions, suspicious API, special string to construct a sample feature space and evaluate the safety of mobile payment applications based on support vector machine algorithm. By comparing the accuracy of a variety of algorithms, the algorithm is turned to be feasible. The importance of the evaluation index is different during the sample classification and there is a difference in the performance of the kernel function under different conditions. So this paper studies the performance of the classifier in different evaluation index set, kernel function and feature weights. It proves that the polynomial kernel support vector machine method is best based on feature weights after the introduction of four types of evaluation index. This method properly avoids the problem of low accuracy due to a single feature while eliminating the adverse effects of weak correlation evaluation index.
BibTeX
@misc{liu2016security,
  author = {Yang Liu and Juan Xu and Hui Fei and Yanhui Guo and Guoai Xu},
  title = {{Security Evaluation Research of Mobile Payment Application Based on SVM}},
  year = {2016},
  doi = {10.2991/icca-16.2016.113},
}
Figure