Resilient by Design: Investigating Backdoor Vulnerabilities in Malware Detection Systems

Resilient by Design: Investigating Backdoor Vulnerabilities in Malware Detection Systems. This report presents a large scale empirical study of backdoor (data poisoning) attacks against ML-based malware detectors. Across 420 experiments spanning poisoning ratios, trigger types and model architectures, the work finds that static malware classifiers exhibit surprisingly strong natural resistance to backdoor attacks, with attack success rates generally below 4%. Tree-based models (LightGBM) show superior robustness compared with neural networks, and simple defences such as ensemble averaging and clean-tuning further reduce attack effectiveness. The study draws on the EMBER feature pipeline and provides code and experimental details to reproduce the results.