Thursday 13 March 2025
The ongoing struggle against malware has led to a cat-and-mouse game between attackers and defenders, with each side constantly trying to outsmart the other. In recent years, researchers have focused on developing machine learning-based Android malware detection (AMD) systems to combat this threat. However, these systems are not immune to attacks, and new methods have emerged to evade detection.
One such method is known as problem space attacks, which manipulate the malicious application files themselves rather than just their features. This type of attack is particularly challenging because it involves modifying the actual malware code, making it harder for detection algorithms to identify.
To combat this threat, a team of researchers has proposed an innovative approach called ADD (Adversarial Defense Framework), designed to enhance the adversarial robustness of ML-based AMD systems against problem space attacks. The framework consists of three main components: space quantification, exponential dimension reduction policy, and pseudo adversarial malware generation.
The first component, space quantification, identifies perturbable features in the feature space that can be modified by an attacker to evade detection. This is achieved through a process called random perturbation, which randomly modifies features within the perturbable space to observe changes in the classification result.
The second component, exponential dimension reduction policy, reduces the dimensionality of the feature space to make it more manageable and efficient for training models. This is crucial because high-dimensional spaces can lead to overfitting and decreased performance.
The third component, pseudo adversarial malware generation, generates a set of pseudo-adversarial malware samples that can be used to train the ML-based AMD system. These samples are generated by randomly perturbing malicious application files to observe changes in their classification results.
The researchers evaluated ADD against three target ML-based AMD systems: Drebin, Drebin-DL, APIGraph, and MaMadroid. The results showed that ADD significantly improved the detection performance of these systems, with an average increase in accuracy of 12.5% across all datasets.
In addition to its effectiveness, ADD is also efficient and scalable, making it a practical solution for real-world deployment. The framework can be easily integrated into existing ML-based AMD systems, providing an additional layer of defense against problem space attacks.
The development of ADD marks an important step in the ongoing battle against malware. As attackers continue to evolve their tactics, researchers must adapt by developing innovative solutions like ADD to stay ahead of the curve.
Cite this article: “Enhancing Adversarial Robustness in Android Malware Detection with ADD Framework”, The Science Archive, 2025.
Malware Detection, Machine Learning, Android Malware, Problem Space Attacks, Adversarial Defense, Feature Space, Dimension Reduction, Pseudo-Adversarial Samples, Classification Accuracy, Robustness.
