Enriching the Assessment of Neural Network-Based Wireless Intrusion Detection Systems: a Case Study on Wi-Fi Networks

Wireless networks are particularly susceptible to cyberattacks due to their open and broadcast-based architecture. Although neural network-based Intrusion Detection Systems (IDS) have shown promising results in identifying malicious activity, their real-world deployment, especially in edge or resource-constrained environments, raises critical concerns. Classification evaluation metrics such as accuracy or f1-score often fail to capture the trade-offs between detection performance, computational complexity, and model confidence. This paper proposes an evaluation framework for IDS in Wi-Fi networks, integrating classical performance metrics as well as assertiveness and computational cost. Building upon an existing Multiple Layer Perceptron (MLP)-based detection method, we analyze how architectural changes impact both detection quality and feasibility of deployment, and subsequently extend our analysis to encompass other state-of-the-art methods. In addition, we introduce a radar-based visualization approach to better represent the multidimensional aspects of model performance, thus showing the importance of balancing accuracy with efficiency and reliability. Finally, we apply a Pareto-based optimization method for trade-off selection based on the radar area and the disparity in metrics.