How Can a Blend of Experiments and AI Make IoT Networks Safer?

Software‑Defined Networking (SDN) is reshaping how we manage connections, especially in IoT devices. To pinpoint which parameters affect performance and security, researchers built a hybrid framework called DOE‑GAN‑SA. The core is Design of Experiments (DOE) with Latin Hypercube Sampling (LHS), ensuring uniform coverage of all possible values. Then a Generative Adversarial Network (GAN) – a type of AI – creates realistic synthetic data, enriching the real dataset.
A classification‑regression tree (CART) checks the quality of the GAN‑generated data. With these tools, an artificial neural network (ANN) performs sensitivity analysis (SA) on a typical SDN‑IoT network. Two scenarios were tested: normal operation and a flooding DDoS attack, measuring throughput, jitter and response time. Results show that merging LHS and GAN improves the detection of critical parameters and boosts DDoS detection accuracy, thanks to synthetic traffic that mimics real flows. In short, blending statistical experiments, AI and sensitivity analysis offers a fresh way to monitor and protect IoT networks.

The DOE‑GAN‑SA framework rests on three methodological pillars. 1) Latin Hypercube Sampling (LHS) is a stratified sampling technique that divides the parameter space into equal intervals, ensuring{ each dimension is represented without overlap. This drastically reduces the number of simulations needed compared to traditional Monte Carlo while preserving high statistical coverage. 2) Generative Adversarial Networks (GANs) consist of two neural networks – a generator and a discriminator – that compete in a zero‑sum game. The generator creates synthetic data from random noise, while the discriminator tries to distinguish real from fake data. Through iterative training, the generator learns the probability distribution of real network data, producing observations that replicate traffic patterns, bursts, and typical DDoS anomalies. 3) The Classification and Regression Tree (CART) acts as a validator: it analyzes the features of synthetic data and checks their consistency with performance metrics (throughput, jitter, response time). If CART detects discrepancies, the GAN training loop is restarted.
Sensitivity analysis (SA) is then performed by an Artificial Neural Network (ANN) which, thanks to non‑linear activation functions, captures complex interactions among parameters such as bandwidth, buffer size, and routing policies. The ANN output provides Sobol' indices that quantify each parameter's influence on the final result. Finally, hypothesis tests (t‑test, ANOVA) confirm the statistical significance of differences observed between normal and DDoS scenarios, ensuring conclusions are not due to chance. This integrated approach reduces predictive model development time, enhances robustness against emerging attacks, and offers a quantitative basis for network policy decisions.