Desain Penilaian Risiko Privasi pada Aplikasi Seluler Melalui Model Machine Learning Berbasis Ensemble Learning dan Multiple Application Attributes

Abstrak

Aplikasi berbasis Android banyak dikembangkan dan tersedia secara bebas di berbagai sumber aplikasi karena sistem operasi Android yang bersifat open-source.  Namun, tidak semua penyedia aplikasi memberikan informasi detail mengenai aspek keamanan aplikasi, sehingga pengguna mengalami kesulitan untuk menilai dan memahami risiko keamanan privasi yang mereka hadapi. Pada penelitian ini kami mengusulkan desain penilaian risiko privasi melalui pendekatan analisis permission dan informasi atribut aplikasi. Kami menggunakan ensemble learning untuk mengatasi kelemahan dari penggunaan model klasifikasi tunggal. Penilaian likelihood dilakukan dengan mengombinasikan prediksi ensemble learning dan informasi multiple application attributes, sementara penilaian severity dilakukan dengan memanfaatkan jumlah dan karakteristik permission. Sebuah matriks risiko dibentuk untuk menghitung nilai risiko privasi aplikasi dan dataset CIC-AndMal2017 digunakan untuk mengevaluasi model ensemble learning dan desain penilaian risiko privasi. Hasil percobaan menunjukkan bahwa penerapan ensemble learning dengan algoritma klasifikasi Decision Tree (DT), K-Nearest Neighbor (KNN), dan Random Forest (RF) memiliki performa model yang lebih baik dibandingkan dengan menggunakan algoritma klasifikasi tunggal, dengan accuracy sebesar 95.2%, nilai precision 93.2%, nilai F1-score 92.4%, dan True Negative Rate (TNR) sebesar 97.6%. Serta, desain penilaian risiko mampu menilai aplikasi secara efektif dan objektif.

Abstract

Since the Android operating system is open-source, many Android-based applications are developed and freely available in app stores. However, not all developers of applications supply detailed information about the app's security aspects, making it difficult for users to assess and understand the risk of privacy breaches they confront. We propose a privacy risk assessment design in this study using an analytical approach to app permissions and attribute information. We use ensemble learning to overcome the drawbacks of using a single classification model. The likelihood assessment is performed by combining ensemble learning predictions and information on multiple application attributes, while the severity assessment is performed by utilizing the number and characteristics of permissions. A risk matrix was created to calculate the value of application privacy risk, and the CIC-AndMal2017 dataset was used to evaluate the ensemble learning model and privacy risk assessment designs. The experimental results show that the application of ensemble learning with the Decision Tree (DT), K-Nearest Neighbor (KNN), and Random Forest (RF) classification algorithms provides better model performance compared to using a single classification algorithm, with an accuracy of 95.2%, a precision value of 93.2%, a F1-score of 92.4%, and a True Negative Rate (TNR) of 97.6%. In addition, the risk assessment design can to assess the application effectively and objectively.