A Biometric-Based Information Security Analysis over Wind Farm Forecasting Using Fractional Learning Model

Abstract

In the context of cybersecurity, detecting anomalies in IoT is a critical concern. In fact, the number of risks and assaults that target smart devices and their interactions is growing along with the prevalence of this technology. Because it is a new method for modelling object interactions and evaluating their accuracy and honesty, behavioral fingerprinting has drawn interest from researchers in this field. However, there are still issues with these AI-based systems' performance.Scalability, privacy, and restrictions on the use of deep learning(DL)approaches are the primary causes. The behavior of a target contact is actually modelled by an object using only the information obtained from direct interactions in classical distributed fingerprinting or image based biometric approaches. This approach provides a narrow perspective on the target, overlooking services and messages exchanged with other neighboring entities. However, there may be serious privacy issues if a target node's behavior like wind farm forecasting application is built into a global model using data from its interactions with its neighbors.To address this problem, the method proposes a novel approach that leverages DL to create a customized global behavioral biometric model for a particular target item, enabling collaborative model training across decentralized nodes by examining how it interacts with various network peers. By using a safe delegation technique to include less capable IoT nodes, our system enables the distributed training of such models. Furthermore, our method ensures the anonymity of the target object and enables secure collaboration among workers by leveraging data privacy and integrity, as well as the strategy's resilience to attacks. A thorough security analysis and a comprehensive experimental campaign are used in this study to assess the resilience, accuracy, and performance of the safe, completely privacy-preserving solution that results from all these aspects. Lastly, our model's performance is quite satisfactory, with an average accuracy score of 99%, reliably differentiating robustly detected anomalies and distinguishing them from normal behaviors across all evaluated test sets.