CNN-Based Real-Time NIDS
Budget: $15 – $25 USD
My final-year project revolves around turning the CICIDS 2018 dataset into a real-time Network Intrusion Detection System that flags anomalies in packet-level traffic. The part I still need is the heart of the system: a well-architected Convolutional Neural Network, trained and tuned so that it can keep pace with live throughput and later be ported to an FPGA.
Here is what I expect to walk away with:
• Clean, reproducible preprocessing of raw pcap files into tensors suited for a CNN, with all scripts delivered.
• A CNN architecture (TensorFlow/Keras or PyTorch) that zeroes in on traffic pattern anomalies rather than signature-based classes.
• Training routines, hyper-parameter sweeps, and clear evaluation metrics (accuracy, recall, F1, ROC-AUC) on the official CICIDS 2018 train/test split.
• Inference benchmark showing latency figures that justify real-time use; please include profiling notes that guide an FPGA hand-off.
• Well-commented code, a short report explaining design choices, and guidance on how the saved model can be converted (e.g., to ONNX or Vitis-AI) for hardware deployment.
Acceptance criteria: the trained model must reach competitive anomaly-detection performance published for CICIDS 2018 and demonstrate <5 ms per packet inference on a standard GPU/CPU test rig.
If you have prior experience compressing CNNs, pruning, or quantizing for FPGA targets, that context will be invaluable, but the core deliverable is the model itself and its training pipeline.
Here is what I expect to walk away with:
• Clean, reproducible preprocessing of raw pcap files into tensors suited for a CNN, with all scripts delivered.
• A CNN architecture (TensorFlow/Keras or PyTorch) that zeroes in on traffic pattern anomalies rather than signature-based classes.
• Training routines, hyper-parameter sweeps, and clear evaluation metrics (accuracy, recall, F1, ROC-AUC) on the official CICIDS 2018 train/test split.
• Inference benchmark showing latency figures that justify real-time use; please include profiling notes that guide an FPGA hand-off.
• Well-commented code, a short report explaining design choices, and guidance on how the saved model can be converted (e.g., to ONNX or Vitis-AI) for hardware deployment.
Acceptance criteria: the trained model must reach competitive anomaly-detection performance published for CICIDS 2018 and demonstrate <5 ms per packet inference on a standard GPU/CPU test rig.
If you have prior experience compressing CNNs, pruning, or quantizing for FPGA targets, that context will be invaluable, but the core deliverable is the model itself and its training pipeline.