Hi, I'm Lewis Birch

I am a Founding Machine Learning Engineer at Mindgard with a First Class Honours degree in MSci Computer Science. I am also currently a PhD student in the Experimental Distributed Systems Labs at Lancaster University, specializing in adversarial machine learning.

I am always eager to expand my knowledge and improve my capabilities through ongoing learning. I am excited to share my work and research with you, and I hope you find it informative and interesting.

Thank you for visiting!

My Experience

Mindgard

Machine Learning
Cyber Security
Team Work
Communication

As a founding machine learning engineer at a university spin-out startup, I am at the forefront of advancing adversarial machine learning security. Leveraging the expertise developed during my PhD at Lancaster University, I focus on creating robust machine learning systems that are secure against adversarial attacks.

In this role, I am responsible for designing and implementing cutting-edge solutions to enhance the security of machine learning models. My work involves the development of algorithms that can detect and mitigate threats from model inversion, membership inference attacks, and evasion techniques. These efforts aim to safeguard sensitive information and ensure the reliability of machine learning applications in real-world scenarios.

I also collaborate closely with a dynamic team of researchers and engineers to integrate these security measures into commercial products and services. Utilizing a variety of technologies, including TensorFlow, PyTorch, and CUDA, we work together to push the boundaries of what is possible in machine learning security, aiming to set new industry standards.

This role not only allows me to apply my technical skills in challenging new contexts but also offers the opportunity to contribute significantly to the field of machine learning security, ensuring safer digital environments for users and businesses alike.

PhD Secure Machine Learning

Research
Machine Learning
Academic Writing
Python

I am currently pursuing a PhD in Secure Machine Learning within the Experimental Distributed Systems (EDS) at Lancaster University. My research focus is on understanding and mitigating the vulnerabilities of the data used to train and develop machine learning models to adversarial attacks.

Specifically, I have been investigating and evaluating various methods such as model inversion, a technique in which an attacker can reverse-engineer the data used to train a model, membership inference, an attack in which the adversary can determine whether a given data point was present within the dataset during training, and model evasion, in which a modified input is generated to attempt to fool a model through misclassification.

In addition to my research, I have been honing my technical skills through the use of various tools and technologies such as Tensorflow, PyTorch, TVM, Docker, CUDA, and machine learning pipelines/operations. These skills are crucial in the advancement of my research and will be beneficial in my future professional endeavors.

MSci Hons Computer Science (with Industrial Experience)

Software Development
Team Work
Industry Experience
Communication

From 2018 to 2022 I attended Lancaster University with an academic scholarship and completed a four-year degree with an integrated master’s in computer science, graduating with a first-class honours classification.

My time at Lancaster has allowed me to develop general and software development-specific skills through multiple collaborative and individual projects. Improving my skillset greatly in areas such as teamwork, leadership, academic research and core software development principles.

During my time at Lancaster, I was elected student representative multiple times by my peers. A role In which I worked to assist the wider faculty academic staff, to improve student experience whilst they undertook the course. This allowed me to develop improved communication and speaking skills, by having to convert student feedback into actionable changes to deliver to course conveyors for implementation.

Barclays Bank

Big Data
Java Spring
Agile
API Development

In 2021 I was selected for a ten-week internship at Barclays Bank PLC, where I was working as a software developer within the wholesale lending cycle. Placed within a team of data engineers at Barclays Technology Centre based at Radbroke Hall in Cheshire.

Within the ten weeks I was involved in developing an API for pulling wholesale lending data into a customisable dashboard for whole lending agents and regulatory oversight officers. This included making use of the Java Spring-Boot framework for web request routing, as well as web development technologies HTML, CSS, and JavaScript for dashboard development.

This opportunity provided me with exposure to real-world software development and deployment cycles within a fast-paced collaborative environment. Giving exposure to the real-world use of methodologies like stakeholder management, agile sprints and unit testing.

Overall, an extremely positive experience, that enabled me to apply software development methods in a real-world context with high-standard peers and leaders. Ultimately resulting in a graduate job offer.

Lancashire Police

Software Development
ElectronJS
Sensitive Data

In 2022 as part of the industrial experience portion of the final year of my master’s in computer science at Lancaster, I was selected to carry out a ten-week placement with Lancashire Constabulary.

For the ten weeks I was working within the Vehicles Investigation Unit (VIU), developing a bespoke desktop application for use as an investigation aide to visualise tracking information during crime-related data exploration. The purpose of the application was to reduce investigation time and effort, increase case throughput and accuracy.

The creation of this application was carried out by using the ElectronJS desktop application development framework, with React being used to build the user interface. Google Map’s JavaScript APIs was used to provide the mapping functionality to filter and overlay data onto different maps. The development process followed the agile methodology, initially starting with user story designs and proceeding to weekly sprints in collaboration with stakeholder feedback from team members of the VIU.

Birch Media Design

Web Development
Project Management
WordPress
Leadership

Since 2018 I’ve been working part-time as a web developer using the open-source WordPress eco-system. I initially took on this role to expand Birch Media Design’s capability from graphic design and video editing, to website development and maintenance. Helping to improve the company’s revenue and enable its clients to expand their online presence.

From this experience, I’ve developed core web development skills in PHP, HTML, CSS, JavaScript, Search Engine Optimization (SEO) and User Experience (UX) design. As well as general communications organisational and communication through liaising with clients on their design requirements, work timelines, expected costs and how to maximise the value of the online presence for their business.

My Publications

Model Leeching: An Extraction Attack Targeting LLMs

Large Language Model
Model Stealing
ChatGPT

In this paper, we introduce Model Leeching, a novel extraction attack specifically designed for Large Language Models (LLMs). This technique effectively distills task-specific knowledge from a target LLM into a more compact model with fewer parameters. Our findings reveal that using Model Leeching on ChatGPT-3.5-Turbo allowed us to achieve a 73% Exact Match (EM) similarity, with SQuAD scores of 75% EM and 87% F1 accuracy, at a minimal cost of only $50 in API expenses.

Furthermore, we explored the potential of using the model obtained through Model Leeching to stage further adversarial attacks on the original LLM. The results demonstrated an 11% increase in the success rate of these attacks when applied back to ChatGPT-3.5-Turbo, showcasing the feasibility and danger of adversarial attack transferability facilitated by this extraction method.

Compilation as a Defense: Enhancing DL Model Attack Robustness

Deep Learning Model
Low-level
Tensor Optimisation

This paper explores an innovative approach to enhancing security in adversarial machine learning (AML) by focusing on side-channel attacks—a critical but often under-investigated area. While previous research acknowledges the severity of these attacks, few efficient solutions have been proposed that do not require extensive and costly model re-engineering. Our study introduces a novel defense mechanism employing model compilation and tensor optimization techniques.

We demonstrate that these methods can reduce the effectiveness of model attacks by up to 43%. The findings also provoke a broader discussion on potential future directions for improving AML defenses without compromising on performance or incurring significant costs.

My Skills

Contact Me

FIll out the form below and I’ll get back to you as soon as possible.