Home/Research

Research

Machine learning, rewritten in qubits.

Four connected threads, each tested against classical methods on real data.

Interactive

How a qubit holds a pixel.

Drag the sliders to rotate one qubit. Then measure it a hundred times and watch probability turn into counts.

|ψ⟩ = 1.00|0⟩ + 0.00|1⟩

P(0)1.00
P(1)0.00

This is the same Y rotation her filters use to load a pixel into a qubit: brighter pixels tilt the state further from |0⟩ towards |1⟩.

The pre-processing filter

Four steps, from image to features.

A simplified walk through the idea behind her quantum pre-processing filter.

  1. 01

    Encode

    Each pixel becomes an angle. A Y rotation gate tilts a qubit by that amount, so brightness is stored in the qubit's state.

  2. 02

    Correlate

    Two CNOT gates link pairs of qubits, so the measured result reflects how neighbouring pixels relate, not just each one alone.

  3. 03

    Measure

    Reading each qubit in the Z basis returns an expectation value between minus one and one. These become the new features.

  4. 04

    Learn

    The quantum features feed a classical neural network, which learns to classify the image as usual.

Quantum generative models

Variational autoencoders with a quantum encoder, built to reconstruct and generate images with higher fidelity.

Her Q-VAE replaces the classical encoder of a variational autoencoder with a quantum down sampling filter. Pixel values are encoded with Y rotation gates and read out through Pauli Z measurement to form the latent representation, while a convolutional decoder rebuilds the image.

Tested on the MNIST and USPS datasets against a classical VAE and a classical direct passing variant, the quantum model reached lower Fréchet Inception Distance scores, a sign of sharper and more faithful reconstructions. The quantum encoding adds no extra trainable parameters, which matters on today's limited hardware.

The same family of models underpins her work on quantum variational autoencoders for time series anomaly detection at CSIRO.

  • MNIST and USPS benchmarks
  • Lower FID than classical baselines
  • No added trainable parameters
  • Open code on GitHub

Quantum pre-processing filters

A compact four qubit circuit that extracts features before a neural network ever sees the image.

The quantum pre-processing filter is deliberately simple: four qubits, Y rotation gates to encode pixel values, and two controlled NOT gates to create correlation between them. Its measured outputs become the features passed into a fully connected neural network.

Across her doctoral papers the filter improved multi-class image classification accuracy, and a follow-up study showed it helping binary classification when only small training samples are available, a common real-world constraint.

  • Four qubits
  • Y rotation encoding
  • Two CNOT gates
  • Published in Sensors, 2023

Quantum era cybersecurity

What quantum computing means for the cryptography that connected devices rely on, and for attacks on AI.

With collaborators she proposed architectures to address public key cryptography weaknesses in the Internet of Things in light of quantum computing supremacy, first at ICTC 2021 and then in an extended Sensors journal article.

Her CSIRO research sat across quantum computation, adversarial machine learning, and cybersecurity and privacy, with anomaly detection as a recurring tool for spotting threats in data.

  • IoT public key cryptography
  • Adversarial machine learning
  • Anomaly detection
  • Australian Women in Security honours

Quantum AI in the real world

Taking quantum machine learning out of the quantum sector and into food systems, schools and speech.

She leads the first CSIRO Industry PhD in quantum machine learning, a four year partnership between Food Ladder, CSIRO and UniSQ. It develops quantum reinforced AI to make urban hydroponic food systems more resource efficient, and builds evidence on how green space exposure supports student wellbeing.

It is one of the first projects in Australia to apply quantum machine learning in a non-quantum industry. Separately she has collaborated on parameterised quantum circuits for speech emotion recognition.

  • Food Ladder partnership
  • Four year Industry PhD
  • Hydroponics optimisation
  • Student wellbeing

Fields of research

Classified by ANZSRC code.

  • 461307Quantum computation
  • 461101Adversarial machine learning
  • 460499Cybersecurity and privacy
A significant step in connecting quantum technologies with real-world industries.
Farina Riaz On the Food Ladder partnership, 2025

Collaborate

Research, speaking, supervision or media.

If it involves quantum AI, cybersecurity or getting more women into both, she would like to hear about it.

Get in touch LinkedIn