Hybrid CNN-LSTM Network Intrusion Detection System Utilizing Multi-Task Learning and Simulated Network Defense

Overview

Cybercrime threatens critical infrastructure, economic assets, and privacy, with financial loss from cybercrime continuing to increase each year. The proliferation of increasingly complex, cross-dependant software, as well as AI-generated code with dubious reliability, calls for improved defense at every level. The proposed deep-learning based Network Intrusion Detection System (NIDS), utilizing a custom hybrid Convolutional-Long Short-Term Memory (LSTM) architecture, solves the inadequacies of previous anomaly-based or signature-based methods, is capable of direct deployment onto real networks, and outperforms contemporary methods. The model utilizes “sub-gates” preceding traditional LSTM gates, which include convolutional layers to extract features from packet inputs and also critically reduce the dimensionality of the input and cell-memory. To improve performance, the model includes an ELU activation function for memory inputs and a modified output gate which bypasses the traditional output selector vector. The model is trained on the CIC-IDS2017 dataset, and outputs per-packet binary classification. Packets are inputted in “streams” (15-60 packets) with a random distribution of malicious to benign packets (10-50%) shuffled in order. The greatest improvement to the model was the inclusion of a small Multi-Layer Perceptron (MLP), which reads cell-memory and estimates the distribution of malicious to benign packets on the current stream, used for training. The multi-tasking training technique yielded significant results, with accuracy going from ∼95.9% to ∼98.3% and recall going from ∼96.3% to >99.8%, outperforming other deep-learning methods within the same dataset. This shows that the cell-memory is better representative of persistent patterns across packets.

Competition history

• CSEF 2026 Computational Science (Senior Division) · Entry S-07-35

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Computational Sciences projects
• CSEF Computational Sciences projects
• Every CSEF 2026 project