Abstract

Background: The rapid growth of the Internet of Things (IoT) and edge computing has brought about major challenges in terms of anomaly detection, especially in the context of IoT and edge computing. Though the accuracy of anomaly detection by the use of deep learning models is high, the computational overhead associated with the execution of the model on the edge device is a major drawback. Moreover, the lack of explainability associated with the model makes the model less usable. Aim: The aim of the proposed study is to develop a unified framework for the development of explainable, lightweight deep learning-based anomaly detection, which balances the trade-off between the accuracy of the model, the efficiency of the model, and the explainability of the model. Methodology: The proposed framework for the development of explainable lightweight deep learning-based anomaly detection includes the integration of a lightweight deep learning model, an adaptive explainability module (XAI), and a dynamic trade-off controller. The proposed framework has been evaluated on benchmark datasets, such as CICIDS2017, NSL-KDD, IoT-23, and lightweight and heavyweight models. Results: The experimental results show that the proposed model has comparable accuracy to the heavyweight models (97.9%) but achieves a significant improvement in latency, energy consumption, and model size. The XAI integration results in a negligible impact on the model's accuracy but causes a latency overhead. However, the proposed adaptive method reduces the latency overhead by up to 35%. The framework is able to achieve the desired operation within the optimal trade-off region between latency and explainability. Conclusion: This study has shown that it is indeed possible to develop secure, efficient, and explainable anomaly detection systems for edge devices by optimizing the trade-off between the two. The proposed framework is a promising solution for real-time IoT security applications. This study opens the doors for many research opportunities in the field of adaptive and resource-aware XAI systems.