Towards IoT device privacy & data integrity through decentralized storage with blockchain and predicting malicious entities by stacked machine learning

Abstract

Blockchain technology offers significant advantages in securing the internet of things (IoT) networks. However, IoT devices remain highly vulnerable to security and privacy threats, making them prime targets for malicious activities. This study addresses key challenges in IoT security, including ensuring device authenticity, preserving data integrity through decentralized storage, and enhancing the explainability of predictive models. To tackle these challenges, a novel approach integrating blockchain and machine learning (ML) is proposed. A stacking-based classification model is introduced to differentiate between legitimate and malicious IoT entities. At the base layer, the model leverages the extra trees, multinomial Naive Bayes, and Bernoulli Naive Bayes classifiers, while the logistic regression with cross-validation classifier functions as the meta-model. The preprocessing pipeline includes data normalization and handling of missing values to improve model robustness. To further strengthen security, a local blockchain is implemented on an IoT device manager to register IoT requestors with unique addresses. The Keccak256 hashing algorithm converts these addresses into hashes, which are securely stored on the local blockchain. The actual data is managed using the interplanetary file system, while block validation is performed using a proof-of-stake consensus mechanism. The proposed model classifies IoT devices with superior performance compared to baseline classifiers. Experimental results demonstrate the effectiveness of the stacking model, achieving notable improvements: a 6.90% increase in macro-recall, a 4.49% improvement in the Matthews correlation coefficient and Cohen’s kappa, a 3.33% enhancement in the macro-F1-score, and approximately a 1.02% gain in accuracy, micro-precision, micro-recall, and area under the receiver operating characteristics curve. Additionally, log loss and Hamming loss are reduced by 50%, indicating enhanced reliability and lower error rates. Results of the proposed stacking model are further assessed using the Friedman statistical test and 10-fold cross-validation techniques. To ensure interpretability, Shapley additive explanations and local interpretable model-agnostic explanations are employed, providing insights into model decisions. These findings underscore the effectiveness of the proposed approach in improving IoT security by combining blockchain for decentralized authentication and explainable ML for transparent decision-making.