Smart IoT Anomaly Detection Using DO-TAO (Dandelion Optimization with T-distribution and Adaptive Opposition) and Personalized Federated Learning

Using the N-BaIoT dataset, this study suggests a DO-TAO-FedPer framework for high-performance anomaly detection in IoT contexts while maintaining privacy. Each IoT device first trains a Variational Autoencoder (VAE) locally, and then uses the Dandelion Optimization with T-distribution and Adaptive Opposition (DO-TAO) algorithm to optimize the VAE's architecture and training hyperparameters. Each device's customized local models are then trained using these ideal hyperparameters. FedPer (Federated Personalization) is then used, in which private decoders are adjusted locally and the shared encoder is aggregated worldwide. This method benefits from collaborative learning while guaranteeing device-specific anomaly detection. According to experimental results, the FedPer global model outperforms standalone local models in terms of generalization and performance, achieving high accuracy (> 99%) across all devices.