Abstract:
Researchers are concerned about the detection of obfuscated Android malware, and multiple studies have been proposed to address certain obfuscation techniques. However, the comprehensive consideration of all obfuscation techniques remains a critical cybersecurity challenge due to their mutations. To tackle this issue, we developed the DeepMalOb approach, which utilizes memory dumping and deep learning with MLP to detect obfuscated malicious applications. Although the approach has yielded satisfactory results, we acknowledge potential security risks associated with MLPs, such as adversarial attacks, model inversion attacks, overfitting, and model biases, which may impact the accuracy and robustness of the MLP model and render it vulnerable to obfuscated malware. To improve the DeepMalOb approach, we propose the use of formal security risk analysis methods with MLP to detect hidden malware in Android by analyzing the security risks associated with the MLP model and the input features used for training.