Adversarial Assembly: Leveraging LLMs to Evade CNN-Based Windows Malware Detection

Problem: A robust Machine learning based malware detection is adversarial data-driven. Generating adversarial examples (AEs), especially for Windows Portable Executable (PE) is challenging due to the functionality preservation, complex structure, and optimal location of the adversarial perturbation. Existing research struggles to find effective obfuscation tools that provide functionality validation and semantically safe perturbation space, causing avoidance of the machine code section due to high sensitivity, even if they represent the primary malware logic. Proposed Method: This research tackles the problem where we used Large Language Models (LLMs)’s code generation capabilities and proposed a zero-shot prompt initiated dead code insertion technique, leading to semantically valid obfuscation while preserving the original program logic. We used the grayscale representation of the .text section’s assembly code as features and assessed their effectiveness across 14 Convolutional Neural Network (CNN) architectures for the malware classification task. For creating adversarial examples, we have used six LLMs (DeepSeek, Codestral, Mistral, Gemini and Nova) and a single zero-shot prompt instructing the dead-code insertion and assessed their obfuscation level using cosine similarity and character-level delta entropy. Results: Our experimental result shows that the .text section represented as grayscale image is equally effective, even outperforming the full binary image representation across the CNN architecture. One of the major findings is that our proposed LLM-based attack achieved a 100% attack success rate (ASR) against different CNN architectures.