CV
🧑💻 GitHub 🗄️ LinkedIn 📙 Google Scholar
Profile 👤
PhD-level ML Security Researcher and Engineer focused on the intersection of adversarial robustness and production-scale deployment. Expert in designing Neural Networks that survive real-world security threats, with a proven track record of translating theoretical research into high-performance systems, including real-time video authentication and large-scale automated analysis pipelines.
Education 🎓
PhD Computer Science, University of Birmingham
September 2019 - February 2025
Completed a PhD in Computer Science with a research focus on developing adversarially robust deep neural networks for image classification. Projects included:
- A graph-based malware detection method with a pipeline that processed 130,000 apps in 18 hours, and processed control-flow graphs consisting of over 50,000 nodes.
- Architecting a novel classification with rejection and recovery method that – in some cases – allowed over 90% of weakly perturbed adversarial inputs to be corrected.
- Creation of a novel technique for analysing network similarity using kernel-based activation maximisation that is dataset independent.
BSc Computer Science, University of Birmingham
September 2016 - July 2019
Graduated with first-class honours (GPA 4.25)
Dissertation: “Evolutionary Methods for Generating Secure Deep Neural Networks” - Evaluated the use of multi-objective evolutionary algorithms to genetically generate neural networks that are more robust to adversarial attacks.
Employment History 💼
Postgraduate Researcher, University of Birmingham & Durham University
February 2025 - July 2026
Led engineering on a method for edit-tolerant video authentication to combat fake media. Coordinated with a multidisciplinary team to translate theoretical cryptographic techniques into practical, efficient methods for video processing capable of signing and verifying videos in real time (signing at 78 frames-per-second, verifying at 109 frames-per-second). This is a UK RISE funded project presented at the 2026 RISE Collaborative Research Showcase & Spring School and (upcoming) CRANE 2026.
Postgraduate Teaching Fellow, University of Birmingham
September 2021 - May 2022
Led marking initiatives by designing and implementing an automated marking system that efficiently evaluated over 600 students’ submissions in under four hours. Leveraged containerisation using Docker for isolated code execution and parallelisation. Worked with lecturers, designing assignments to align with learning objectives and assessment criteria. The system enabled the use of practical assignments for assessment and ensured consistent and timely feedback for students.
Postgraduate Teaching Assistant, University of Birmingham
September 2019 - September 2021
Facilitated the delivery of modules by leading interactive learning sessions and helping students grasp complex concepts and develop practical skills. Collaborated closely with faculty members and fellow TAs to create teaching materials, design assignments, and assess student progress, ensuring alignment with learning objectives.
Publications 📃
Using Reed-Muller Codes for Classification with Rejection and Recovery
Daniel Fentham, David Parker, Mark Ryan (2023)
Developed a novel ensemble-based classification-with-rejection and recovery method inspired by error correcting codes to improve robustness of AI models against adversarial inputs.
- Designed a new architecture (RMAggNet) integrating Reed-Muller error-correcting codes to enable both rejection and correction of adversarial inputs.
- Demonstrated recovery of the true class for the majority of low-perturbation adversarial examples, outperforming existing rejection-based defences.
- Showed improved robustness under adversarial training while introducing fewer natural adversaries compared to baseline methods (Thesis work).
- Proved that ensemble size scales logarithmically with the number of classes, enabling efficient deployment for large classification tasks.
Presented at Foundations & Practice of Security 2023 - Published in LNCS
ARASH: Video Authentication from Redactable Hash (in progress 🏗️)
Xiao Yang, Daniel Fentham, Shize Deng, David Oswald, Mark Ryan
Engineered a novel method for authenticating videos while enabling privacy-preserving edits.
- Translated the theoretical cryptographic protocol into an efficient working system.
- Implemented GPU-accelerated video processing using JAX and PyTorch, signing a 2-minute 720p video in less than 40 seconds.
- Evaluated our method against real world datasets, providing guarantees of robustness and sensitivity even in adversarial scenarios.
- Designed experiments to evaluate the efficacy of ARASH including the generation and augmentation of bespoke datasets.
Projects 🛠️
AndroCFG
Developed a scalable pipeline to construct control flow and function call graphs from disassembled Android applications for downstream ML-based malware detection.
- Reverse-engineered and parsed Smali bytecode to generate CFGs, function call graphs, and hybrid graph representations.
- Implemented function-expansion techniques around suspicious call sites to enhance structural context for classification.
- Added compatibility with established malware detection frameworks (e.g., CFGExplainer, MalGraph).
- Added
dotexport for graph visualisation and analysis. - Optimised for high-throughput processing within a toolchain handling 130,000 APKs (decompilation, graph generation, and feature extraction) on the SLURM high-performance compute platform.
Pokémon RL (in progress 🏗️)
Developing a reinforcement learning agent capable of autonomously playing and winning turn-based battles in Pokémon Crystal by interfacing directly with emulator memory.
- Reverse-engineered Game Boy Color RAM to construct a controllable RL training environment.
- Programmatically manipulated in-memory game state (Pokémon parties, levels, stats, moves) to generate diverse training scenarios.
- Designed abstract state representations to reduce dimensionality and improve training efficiency.
- Implementing and evaluating RL policies for competitive battle performance.
Letsfool AI
Developed and deployed an interactive AI application enabling users to manipulate handwritten digit inputs to fool a classifier using model explainability feedback.
- Built and deployed a cloud-hosted web application using Google Cloud Run.
- Containerised the service with Docker to enable portability and scalable deployment.
- Implemented CI/CD pipeline from GitHub for automated testing and production updates.
- Enhanced baseline MNIST classifier with robustness improvements to better handle user inputs.
- Integrated explainability methods to visualise model sensitivity and guide user-driven adversarial exploration.
VGSum
Designed and implemented an autonomous LLM-based system to ingest, contextualise, and summarise video game news from Bluesky.
- Integrated with the ATProtocol to extract posts, linked articles, and metadata from curated video game news accounts.
- Built an agentic workflow using LangChain and LangGraph to summarise inputs and add context using tools.
- Implemented external tool integration to retrieve OpenCritic scores using RapidAPI, incorporating review data into summaries.
- Evaluated agent behaviour and performance using LangSmith, identifying latency and tool-calling bottlenecks.