Enhanced Intrusion Detection for Cloud-Integrated IoT Networks Using Autoencoder-Driven Unsupervised Deep Learning

With the increasing integration of IoT devices into cloud environments, robust Network Intrusion Detection (NID) systems are critical for ensuring data security and network resilience within these interconnected infrastructures against sophisticated cyber threats. Conventional NID methods often struggle to keep pace with the dynamic and complex traffic patterns in cloud-connected IoT networks, necessitating advanced deep learning techniques for effective threat detection and anomaly identification. This research provides a comprehensive performance and convergence analysis of unsupervised autoencoder models, including dense autoencoder, variational autoencoder (VAE), long short-term memory autoencoder (LSTM-AE), and convolutional neural network autoencoder (CNN-AE), for NID tasks. We experiment with a standard dataset to analyze the convergence properties by evaluating the behavior of the reconstruction loss function L(θ). We evaluate the anomaly detection efficacy of each model utilizing F1 scores and accuracy. The results show that the VAE is consistently better than other models for numerical stability. It has the smoothest, most monotonic de-crease in L(θ), the least variation in the gradient norm ∇L(θ), and the fastest convergence. Furthermore, the VAE attains the highest F1 score of 89.64%, closely followed by the LSTM-AE at 86.50%. The CNN-AE and dense autoencoder models, albeit successful, have slower convergence and marginally reduced detection accuracy. This thorough research highlights the trade-offs between convergence behavior and detection performance, establishing the VAE as the most resilient and effective model for real-time NID applications, especially in settings requiring strong numerical stability and swift model convergence.