Building a Security System Using a Deep Learning and Blockchain-Based Approach for Abnormal Event Detection and Data Security

Abstract

Building surveillance systems in smart cities face significant challenges in ensuring security, detecting anomalies, and processing large volumes of real-time data. This paper proposes an enhanced surveillance system that integrates lightweight deep learning models with blockchain technology to improve authentication and authorization. The system processes real-time data from surveillance cameras, using a dataset of 46,297 samples divided into training, validation, and testing subsets. It operates in two main stages. In the authentication stage, images from surveillance footage are processed by three lightweight CNN models to predict class values, which are then hashed using the SHA-256 algorithm. These hashes are structured into a Merkle tree and stored in a blockchain network. In the authorization stage, the system verifies whether a user has blockchain-registered access to surveillance data for a specific time period. If authorized, the system retrieves data from all cameras and uses a SoftMax function to classify activity as normal or abnormal. Among the CNN models tested, DenseNet201 showed superior performance with high accuracy and low loss in training and validation. In testing, DenseNet201 achieved the best results with a precision of 0.951, recall of 0.946, F1 score of 0.948, and overall accuracy of 0.936. These findings highlight the model’s effectiveness in identifying abnormal behavior and the potential of integrating deep learning with blockchain to secure and manage smart building surveillance systems.