A lightweight, RV32IMF RISC-V soft-core ("Hornet") implemented on FPGA for edge-based network intrusion detection using a custom C-based MLP inference engine.

• Design and Implementation of RISC-V-Based Core For AI Applications
  • 📖 Table of Contents
  • 🏛 System Architecture
  • 🚀 About the Project
  • 🏗 System Design & Verification
    • 1. RTL Simulation Flow
    • 2. Hardware-in-the-Loop (HIL) Flow
  • 📂 Repository Structure
  • 🛠 Prerequisites & Toolchain
    • Hardware Requirements:
    • Python Dependencies
  • ⚙️ How to Run (Hardware Verification)
  • 👥 Team & Acknowledgments
    • Authors
    • Acknowledgments
  • 📬 Contact

The design features the Hornet core (RV32IMF) integrated with a custom Wishbone wrapper and peripherals on the Nexys Video FPGA.

This project presents the design and implementation of a lightweight, RISC-V-based Edge AI processor capable of executing deep learning inference for network security applications. The primary objective is to validate the feasibility of running a complex, floating-point-heavy Multi-Layer Perceptron (MLP) on a soft-core processor within an FPGA environment.

Key Features:

• Core: "Hornet" Processor (RV32IMF) – A custom 32-bit RISC-V core with Single-Precision Floating-Point support.
• Hardware: Implemented on Xilinx Nexys Video (Artix-7 XC7A200T).
• Software: Custom C inference engine (no heavy frameworks like TFLite) built from scratch.
• Algorithm: Custom Focal Loss MLP trained on the NSL-KDD dataset for 5-class network traffic classification.
• Performance: Achieved 78.87% accuracy on hardware, matching the software model.

Before going into the HIL flow, the design was tested on RTL Simulation using Vivado & XSim. The first verification topology is:

To ensure reliability, we developed a comprehensive verification suite. This flow validates the design by cross-referencing the Python Golden Model against both the RTL Simulation and the Physical FPGA output via UART, ensuring bit-accurate inference results. The second verification topology is:

This repository is organized by version and functionality.

.
├── v_1.0_stable/            # PRIMARY STABLE RELEASE
│   ├── no_batch/            # Sequential inference implementation (No-Batch processing)
│   │   ├── MLP_No_Batch     # Main Vivado Project & RTL Source
│   │   ├── Python_Codes     # Training scripts, dataset pre-processing, and model weights
│   │   └── test/            # Simulation and Hardware verification testbenches
│   │
│   └── source/              # Core source files (RTL, linkers, bootloaders)
│
│
└── README.md                # Top-level documentation

This project relies on specific tool versions to ensure synthesis reproducibility and accurate behavioral simulation.

• Development Board: Digilent Nexys Video (Artix-7 FPGA)
• Chip Compatibility: Validated on XC7A200T (Artix-7) and XC7A100T (Artix-7).

Required for training the model and running the UART test suite:

pip install tensorflow==2.20.0 numpy==2.2.6 pandas==2.3.3 scikit-learn==1.7.2 pyserial==3.5 matplotlib==3.10.7

1. Synthesize & Bitstream: Open the project located in v_1.0_stable/no_batch/MLP_No_Batch using Vivado. Run Synthesis, Implementation, and Generate Bitstream.

2. Program FPGA: Connect your Nexys Video board via USB-JTAG and program the device with the generated .bit file using the Hardware Manager.

3. Run UART Listener: On your host machine, navigate to the Python scripts directory and start the serial listener to capture inference results.

   cd v_1.0_stable/no_batch/Python_Codes
   # Replace '/dev/ttyUSB1' with your actual COM port (e.g., COM3 on Windows)
   python3 serial_listener.py --port /dev/ttyUSB1 --baud 115200

4. Trigger Inference: Press the CPU Reset button on the FPGA board to initialize the Hornet core and begin the classification loop. The results will appear in your terminal.

This project was developed at Istanbul Technical University (ITU) as a graduation thesis.

• Yusuf Tekin (tekiny20@itu.edu.tr)
  • Role: Software Design, Hardware/RTL Design, System Verification.
• Ahmet Tolga Özkan (ozkanah19@itu.edu.tr)
  • Role: RTL Design for Hardware Accelerator.

We would like to express our deepest gratitude to our advisor, Prof. Dr. Sıddıka Berna Örs Yalçın, for their invaluable guidance, technical insights, and mentorship throughout the development of the "Hornet" core.

For inquiries regarding the hardware architecture, the software inference engine, or potential collaborations, please contact the authors via email or open an issue in this repository.

• Yusuf Tekin: tekiny20@itu.edu.tr
• Ahmet Tolga Özkan: ozkanah19@itu.edu.tr