LLMalMorph: On The Feasibility of Generating Variant Malware... | AI Research

Описание: Discussion of the paper 'LLMalMorph: On The Feasibility of Generating Variant Malware using
Large-Language-Models'.
LLMalMorph explores the feasibility of using Large Language Models (LLMs) to generate malware variants from source code, addressing the increasing threat of AI-powered cyberattacks. The core problem is to leverage pre-trained LLMs, without resource-intensive fine-tuning, to create functional malware variants capable of evading antivirus engines and Machine Learning (ML) classifiers, while preserving original malware semantics.

The paper introduces **LLMalMorph**, a semi-automated framework designed for Windows malware written in C/C++. Its methodology involves two main modules:
1. *Function Mutator:* Extracts function-level information (function bodies, headers, global variables) from malware source code. It then uses custom-engineered prompts and strategically defined code transformations to guide an open-source LLM (Codestral-22B) in modifying these functions.
2. *Variant Synthesizer:* Integrates the LLM-transformed functions back into the source code. A crucial "human-in-the-loop" process is incorporated for debugging and resolving compilation errors, particularly for multi-file malware projects and complex transformations.

LLMalMorph employs six code transformation strategies:
*Code Optimization:* Removes redundancies, fixes bottlenecks, simplifies logic.
*Code Quality and Reliability:* Improves error handling, adheres to coding standards.
*Code Reusability:* Splits large functions into smaller, modular blocks.
*Code Security:* Addresses vu...
URL: https://huggingface.co/papers/2507.09411

#AI #MachineLearning