Shipped: Quantinuum Hybrid AI — Compounded Adversarial Attack Research

An open-source adversarial AI research proof of concept exploring how hybrid quantum-classical neural networks respond to white-box, targeted, and compounded adversarial attacks.

The framework combines classical neural network processing with quantum circuit operations implemented using the Quantinuum pytket and pytket-qiskit libraries.

Using the MNIST handwritten digit dataset, the experiments explore how multiple adversarial techniques can be combined to manipulate classification outcomes.

The framework integrates TensorFlow, PyTorch, and TorchAttacks to study adversarial perturbations, model predictions, and classification performance before and after attacks.

The project includes executable research code and visualizations illustrating changes in hybrid neural network behavior under adversarial conditions.

Its objective is to better understand vulnerabilities in emerging quantum-classical machine learning architectures and provide a foundation for future defensive research.

The project evaluates an undefended hybrid model and should not be interpreted as demonstrating quantum resistance to attacks or a general quantum advantage.