Deep Learning based Feature Fusion Network for Long Range Attack Detection on Blockchain Consensus Layer

Abstract

A blockchain is recognized as a revolutionary and advanced technology, primarily due to its features of privacy, security, immutability, and data integrity. The consensus layer serves as the foundation and it is the most critical component of blockchain architecture. Identifying Long-Range Attacks (LRA) within a block chain presents significant challenges. Existing studies face various difficulties in detecting these long-range attacks and monitoring the behaviour of validator nodes within the blockchain network. Consequently, this paper introduces a novel deep learning approach designed to accurately detect the nodes as either normal or attack, thereby effectively reducing the risk of long-range attacks. Initially, data are collected, and pre-processing is done to improve the quality of input using Upgraded Min-Max Normalization. Next, high level features are extracted using Improved Non-negative Matrix Factorization (INMF) and Sparse Variational Auto encoder (SVAE) methods. The INMF based features are given as input to the Depth wise Separable Convolutional Resnet (DSC-ResNet) to learn the latent features. The Stacked Bidirectional Gated Recurrent Dropout Network (SBi-GRDN) is trained using the SVAE features to identify complex relationships and interactions among features for capturing attack patterns efficiently. Then, the attention layer is used to fuse features from the DSC-ResNet and SBi-GRDN models. Finally, a fully connected layer with a sigmoid is employed for classifying the attack. The experimental results illustrate that the proposed approach attains an accuracy of 99.1%, Precision of 99.5%, Recall of 99.4%, and F1-score of 99.5%, which provides effective results in detecting long range attacks.