Finding Biases and Backdoors in an NVIDIA Morpheus Spam Detection Model

Pretrained models have become an essential tool for various industries and applications, including for personal use, and recently, cybersecurity. The use of pretrained models has significantly reduced the time and resources needed to develop and deploy machine learning models as they only need to be fine-tuned for specific use cases. However, as pretrained models become more widespread, it has become evident that these models may be vulnerable to various types of attacks, including biases, backdoors, and other potential vulnerabilities. Furthermore, the vulnerabilities that are harbored in pretrained models may be passed on to the fine-tuned versions of these models. This research paper aims to explore these vulnerabilities, especially as they relate to cybersecurity. Additionally, this paper focuses on, and finds, vulnerabilities in a NVIDIA Morpheus pretrained model.