Abstract

Android smartphones have been a popular means for people to connect to digital services and hence a lucrative target for malicious attacks. The Portable Document Format has also been a popular means of sharing files on Android platforms and has also started to be used increasingly for malicious attacks through inbuilt scripts, objects, and obfuscation methods in PDF files. Signature-based methods have been inefficient in defending against modern and novel attacks. A machine learning-based approach for malicious PDF file detection in an Android Security Advisor application is described in this paper. The developed system extracts structural and metadata-based features from PDF files and utilizes a neural network-based model trained offline and implemented on Android smartphones via TensorFlow Lite. The detection framework preserves parity between training and mobile-based data preprocessing via metadata-based normalization. The experimental outcome indicates that the developed technique is highly efficient in malicious PDF file detection and has a lower false positive ratio and is optimized for Android-based platforms.