Mathematical modelling of ensemble-based federated learning framework for false data injection attacks detection in artificial intelligence of things

Abstract Rapid evolution of Artificial Intelligence of Things (AIoT) has allowed intelligent data-driven decision-making under several cyber-physical systems (CPS) like smart cities, smart grids, industrial automation, and healthcare monitoring. AIoT incorporates artificial intelligence (AI) methods with IoT devices to develop large volumes of sensor data and improve system efficacy, automation, and predictive proficiencies. Among different forms of attacks, the false data injection attack (FDIA) has proven that high-priority cyber-related problem and has received more attention recently. The FDIA detection is an important security process in the smart grid controller system and is resolved by data-driven machine learning (ML) algorithms. Now, Federated Learning (FL) is a well-known ML method that assists in addressing data privacy problems by permitting users to collect its raw data locally and transferring it local model parameters only to the aggregator server to collectively train a shared global model. This paper develops a Federated Learning-Based Framework for Detecting Attacks Using an Ensemble Classification Model (FLFDA-ECM) in distributed cybersecurity systems within the Artificial Intelligence of Things (AIoT) environment. The major intention of the FLFDA-ECM paper is to propose an FL-based cybersecurity framework for detecting and classifying FDIA in AIoT- enabled distributed systems while preserving data privacy. At first, the data pre-processing contains dual levels, such as handling missing data and feature scaling to regularize the input data. In the feature selection stage, the hybrid approach integrates minimum redundancy maximum relevance and ReliefF techniques to recognize the most appropriate attributes. Furthermore, an ensemble-based weighted majority voting approach incorporating bi-directional gated recurrent unit, variational autoencoder, and temporal convolutional network models is employed for attack detection. Finally, the parameter fine-tuned model has been executed by the Adan optimizer to enhance the classifier solution. The performance evaluation of the FLFDA-ECM technique occurs using a benchmark dataset. The empirical outcomes specified the boosted performance of the FLFDA-ECM model in comparison with recent techniques.