The proliferation of Internet of Things (IoT) enabled cyber-physical systems, including industrial equipment and operational IT, facilitates the transmission and reception of data over the internet. These systems are equipped with sensors that monitor equipment conditions and report findings to a centralized server. However, malicious users can potentially compromise these sensors, altering the data they transmit. This falsified information can lead to erroneous actions being taken based on inaccurate data, resulting in equipment failures and production disruptions in various sectors. Despite the development of numerous algorithms aimed at detecting such cyber-attacks, many of these approaches struggle with the issue of data imbalance, where one class (e.g., normal records) significantly outnumbers the other class (e.g., attack records). This imbalance often hampers the accuracy of detection algorithms. Existing solutions typically rely on over-sampling or under-sampling methods to generate synthetic records for the minority class, which can introduce biases. To address the limitations of these traditional techniques, we propose a novel method that does not rely on under-sampling or over-sampling. Our approach consists of two main components, which we will outline in the subsequent sections.