LAMLAD: LLM-Based Adversarial Attack Against Machine Learning for Android Malware Detection

The increasing sophistication and volume of Android malware have driven the adoption of Machine Learning (ML) models for scalable and accurate threat detection. However, these models remain vulnerable to adversarial attacks that subtly manipulate input features to evade classification. In this paper, we introduce LAMLAD, a novel adversarial attack framework that exploits the generative and reasoning capabilities of Large Language Models (LLMs) to bypass ML-based malware detectors. LAMLAD employs a two-agent architecture composed of an LLM manipulator, which crafts realistic feature-level perturbations without altering core malicious behaviors, and an LLM analyzer, which guides the modification process to ensure successful evasion. To improve efficiency and context awareness, LAMLAD integrates Retrieval-Augmented Generation (RAG) into the LLM workflow. By targeting Drebin features, LAMLAD enables stealthy and high-confidence attacks against widely used Android malware classifiers. We evaluate LAMLAD against three representative ML-based Android malware detectors and compare it with two state-of-the-art adversarial attack techniques. Results demonstrate that LAMLAD achieves an attack success rate of 97% with an average of 3 attempts per adversarial example, highlighting its potency, efficiency, and adaptability in real-world adversarial scenarios.