<p>Blockchain networks are still quite susceptible to disruptive network-layer attacks, even if they provide decentralization and resistance to tampering. The high dimensionality, dynamic patterns, and adversarial noise of blockchain communication frequently prevent traditional intrusion detection algorithms from generalizing to it. The proposed model suggests TA-FRCNet, a Twin Attention-based Fine-tuned Residual Capsule Network, as a solution to these problems. It combines a residual capsule backbone with dual-channel spatial attention modules to improve discriminative feature learning from intricate blockchain traffic. Localized anomalies and global traffic correlations are both captured by the twin attention mechanism, while residual connections strengthen deep capsule routing and lessen vanishing gradient problems. Using the Honey Badger Optimization technique, hyperparameters are adjusted to further improve detection performance. This model analyzes data in key components: collecting data, preprocessing traffic, extracting relevant features, and finally, detecting the attack. In this method, critical components were extracted using Squeeze Excited EfficientNetB0. Experimental results demonstrate that the proposed model achieved an accuracy of 98.77% on the SNMP2016 dataset and 98.62% on the CSE-CIC-ID2018 dataset, and the accuracy of the BNat dataset is 98.5%, thereby providing a robust solution for contemporary blockchain network challenges.</p>