Research on Deep learning Algorithm for Intrusion Detection based on Autoencoder

In light of the growing prevalence of cyber security threats, the necessity for effective intrusion detection has become paramount. In order to address the limitations of existing intrusion detection algorithms, particularly their underperformance and failure to capture feature differences in small datasets, this paper proposes an advanced autoencoder-based intrusion detection algorithm, SSAE-ResNet. The algorithm incorporates the feature learning capabilities of a Stacked Sparse Autoencoder (SSAE) to reconstruct network flow data, compute feature residuals, and dynamically adjust prediction thresholds in order to circumvent overfitting. Furthermore, a Residual Network (ResNet) is utilised to accurately identify key patterns and anomalies. To address the issue of data imbalance, a hybrid sampling strategy is introduced. The experimental results demonstrate that the SSAE-ResNet algorithm achieves a high recall of 99.63% and a precision of 98.34%. Compared to using only reconstructed features, the recall and precision improved by 6.64% and 4.61%, respectively. The proposed algorithm exhibits superior performance in key metrics and effectively addresses contemporary challenges in intrusion detection.