Deep-NIDS

A deep-learning intrusion detection system designed for localized, real-time packet analysis: traffic features are reduced with PCA, an LSTM autoencoder learns to reconstruct normal traffic, and anomalies are flagged by reconstruction error rather than a fixed signature set.

Raw NSL-KDD traffic features are extracted and normalized, reduced from 41 to 10 dimensions via PCA, then fed through an LSTM autoencoder whose reconstruction error over the input sequence is thresholded to classify traffic as normal or anomalous.

Built on PyTorch with CUDA acceleration for the inference engine, and NumPy/Pandas for the preprocessing pipeline that cleans and vectorizes raw network features into normalized tensors ahead of inference.

Achieves 85.4% detection accuracy on the NSL-KDD evaluation set after reducing the feature space from 41 to 10 dimensions, with 12 ms inference latency per sample — low enough for line-rate, edge-local deployment without a cloud round trip.

Balancing the dimensionality reduction aggressive enough to keep inference latency low against retaining enough signal in the reduced feature set for the autoencoder to separate normal from anomalous traffic reliably.