DeFeAt: Explanable TinyML-based Framework for Anomaly Detection in Nano-Sensor Traffic in 6G-Enabled Smart Grid

This paper presents an Explainable Artificial Intelligence (XAI) based methodology for anomaly detection framework, i.e., DeFeAt, to detect anomalies in nano-sensor traffic in smart grids. TinyML is also incorporated in the framework besides other deep learning models like LSTM and GRU, along with optimizers like Adam, Nadam, and RMSprop. TinyML surpassed the other models as it maintained an average validation accuracy of 98.12 % for all optimizers with the minimum validation loss of 0.3343 using the Adam optimizer and has a model size of 56.26 KB. GRU and LSTM also delivered similar results, however, they needed larger model capacities – with GRU requiring 237.73 KB (Nadam) and LSTM 293.85 KB (Nadam). To promote accessibility, SHAP was used to explain the given model’s prediction, and LIME was also used, the major insights indicated that latency and error rate were the primary emerging features. The proposed DeFeAt framework offers costeffective, accurate, and interpretable fault detection which opens up possibilities for next-generation work on scalability, real-time and online decision-making, and integration with other nextgeneration technologies such as edge computing.