About Me

Profile Image

Hi, I’m Matt, a PhD student at the University of Warwick’s Mathematics for Real-World Systems CDT. My research focuses on making machine learning systems more robust, secure, and understandable. I’m driven by a curiosity for understanding how and why deep learning models work, from their core training dynamics to their behaviour in complex, real-world scenarios.

Some of my recent projects and interests include:

  • Backdoor and weight manipulation attacks in facial recognition systems
  • Reconstructing training data from audio classifiers
  • Training dynamics of neural networks

Selected Projects

Variance Dichotomy in Feature Spaces of Facial Recognition Systems

  • Published in Transactions on Machine Learning Research (TMLR) (Bowditch et al. 2025).
  • Demonstrates that several open-source facial recognition networks (e.g. FaceNet, ArcFace, AdaFace) exhibit a “variance dichotomy” i.e. their learned feature vectors approximately lie in a low-dimensional subspace of the full embedding space.
  • Reveals that this structure acts as a weak intrinsic defense against sequentially installed weight manipulation (backdoor) attacks proposed by (Zehavi, Nitzan, and Shamir 2024), where adversaries attempt to anonymise specific identities by directly modifying model parameters without retraining.

Accuracy of ArcFace under sequential weight manipulation attacks. Benign accuracy initially drops sharply, then unexpectedly recovers (“double descent”), revealing a structural resistance to backdoor attacks.
  • Explores how attackers can circumvent this defense, proposing a modified approach that allows a far greater number of successful backdoors to be installed while maintaining benign accuracy.

A modified backdoor method avoids the sharp drop in benign accuracy, enabling many more successful backdoors.

Reconstructing Training Data from Audio Classifiers with Diffusion Based Enhancement (Ongoing)

  • While generalising effectively to new data, neural networks can retain information from their training examples as a side effect of training.
  • This information can be extracted from the parameters of trained models, potentially leaking sensitive data.
  • This project use investigates the extent to which training data can be reconstructed from audio classifiers, and whether diffusion models can be used to enhance the quality of these reconstructions.

Original training data

Reconstruction from classifier (with diffusion enhancement)

Education

  • PhD in Mathematics for Real-World Systems (2023 - Present)
    University of Warwick, UK

  • MSc in Mathematics for Real-World Systems (Distinction) (2022 - 2023)
    University of Warwick, UK

  • MMath in Mathematics (First Class Honours) (2017 - 2021)
    University of Bath, UK

Work Experience

  • Customer Data Analyst (2021- 2022)
    Screwfix Head Office, Yeovil, UK

Contact

Please feel free to contact me via email: [email protected]

You can also find me on: GitHub

Publications

Bowditch, Matthew, Mike Paterson, Matthias Englert, and Ranko Lazic. 2025. “Variance Dichotomy in Feature Spaces of Facial Recognition Systems Is a Weak Defense Against Simple Weight Manipulation Attacks.” Transactions on Machine Learning Research.

References

Zehavi, Irad, Roee Nitzan, and Adi Shamir. 2024. “Facial Misrecognition Systems: Simple Weight Manipulations Force DNNs to Err Only on Specific Persons.” CoRR.