The growing prevalence of cyberattacks, especially on network-connected systems, highlights the importance of accurate intrusion and attack detection (AD) models. Many machine learning approaches achieve high accuracy, yet their performance depends strongly on internal parameters and dataset choice, which are often underexplored. This paper investigates an Auto-Associative Dense Random Neural Network (AADRNN) for AD, which has been successfully proposed for several applications. We examine in-depth the sensitivity of its performance to two key parameters: the decision threshold (γ) and the number of successive packets used to compute the traffic metrics (I). The model is evaluated on two datasets: the Mirai Botnet dataset and synthetic flood attacks generated on an experimental IoT testbed. Results show how parameters tuning affects Accuracy, True Positive Rate (TPR), and True Negative Rate (TNR), offering insights for designing efficient, parameter-aware AD systems.