ML Transferability for Malware Detection: Cross-Domain Generalization Failures Under Obfuscation
arXiv·medium signal
Study reveals that ML malware detection models trained on clean samples fail to transfer when adversaries use obfuscation techniques, quantifying the transferability gap across detection approaches. The paper systematically evaluates which architectures maintain detection accuracy under evasion and which collapse. Practical implications for any org deploying ML-based security tooling in production.