Projects

Spiking Neural Network Model of Hippocampal Reverse Replay

Dhruti Davey, Eleni Vasilaki

M.Sc. Dissertation – University of Sheffield, 2024–2025

  • Converted a rate-based hippocampal reverse-replay Reinforcement Learning model into a Spiking Neural Network, encoding spatial position via Poisson-firing place cells and modelling action selection with LIF neurons.
  • Simulated a navigation agent performing biological reverse trajectory replay, mirroring hippocampal memory consolidation mechanisms to accelerate reinforcement learning.
  • Validated that the spiking model preserved task performance equivalent to the original rate-based model, while significantly increasing biological plausibility of the architecture.

[PDF] [Code]

Neuromorphic Algorithms Spiking Neural Networks Reinforcement Learning Hippocampal Replay Leaky Integrate-and-Fire Poisson Spiking Neurons Place Cells Spatial Navigation Biologically Plausible AI

Publications

Large Language Model-Based Intrusion Detection System

Dhruti Davey, Kayvan Karim, Hani Ragab Hassen, Hadj Batatia

Springer Nature (published in a peer-reviewed volume), 2024

  • Developed an LLM-based intrusion detection system that classifies network traffic as benign or malicious.
  • Analyzed NetFlow network data and logs, and fine-tuned large language models for accurate classification.
  • Processed structured network traffic data to identify and flag potential security threats.

[PDF] [Code]

cybersecurity large language models intrusion detection network security

M.Sc. Projects

Scalable Machine Learning with PySpark on HPC

  • Processed large-scale datasets on HPC using PySpark, mining NASA web logs with distributed aggregations.
  • Tuned Logistic Regression, Poisson Regression, & ALS collaborative filtering via cross-validated grid search.
  • Compared Random Forest, GBT, and Neural Networks across training sizes, analysing accuracy and latency.

PySpark Scalable ML Distributed Computing Log Mining Recommender Systems ALS Logistic Regression Gradient Boosting Random Forest Neural Networks K-Means Clustering Hyperparameter Tuning Cross-Validation Data Visualisation

Feedforward Neural Network for News Topic Classification

  • Built a Feedforward Neural Network from scratch in NumPy, implementing backpropagation,ReLU, softmax, cross-entropy loss, and dropout for 3-class news article classification.
  • Benchmarked randomly initialised embeddings against frozen GloVe pre-trained embeddings, achieving up to 87.9% test accuracy via SGD with systematic hyperparameter tuning.
  • Extended the network with multi-layer architectures and conducted error analysis using confusion matrices and per-class word frequency inspection to diagnose misclassifications.

Neural Networks NLP NumPy Feedforward Neural Network Backpropagation SGD GloVe Embeddings Text Classification Dropout Regularisation Hyperparameter Tuning Softmax Cross-Entropy Loss AG News Corpus Deep Learning from Scratch

Unsupervised Learning and Deep Classification on Fashion MNIST

  • Applied K-Means clustering and PCA (3-component dimensionality reduction) to Fashion MNIST, visualising cluster quality against true labels in reduced feature space.
  • Trained and benchmarked Logistic Regression, RNN, CNN, and a Fully Connected CNN in PyTorch, evaluating training, validation, and test accuracy across all architectures.
  • Analysed complexity-performance trade-offs across models, implementing structured training loops with cross-entropy loss and batch-based data loading throughout.

Unsupervised Learning Deep Learning PyTorch CNN RNN Logistic Regression K-Means Clustering PCA Dimensionality Reduction Image Classification Fashion MNIST Cross-Entropy Loss Model Comparison Computer Vision

Probabilistic Bee Path Tracking with Gaussian Basis Regression

  • Implemented a probabilistic regression model using Gaussian Radial Basis Functions to reconstruct bee flight trajectories from noisy 2D sensor observations.
  • Derived and minimised a regularised negative log-likelihood objective, fitting 20 model parameters via numerical optimisation using scipy.optimize.minimize.
  • Selected optimal regularisation and noise scale hyperparameters through grid search across multiple datasets, evaluating reconstruction quality against ground-truth flight paths.

Gaussian Radial Basis Functions Probabilistic Modelling Regression Numerical Optimisation Regularisation Hyperparameter Tuning Trajectory Reconstruction NumPy Log-Likelihood Machine Learning Spatial Modelling

Denoising Autoencoder on CIFAR-10

  • Built a convolutional denoising autoencoder in PyTorch using Conv2d and ConvTranspose2d layers to learn compressed latent representations of CIFAR-10 images.
  • Trained the model to reconstruct clean images from synthetically corrupted inputs, evaluating reconstruction error on a held-out noisy test set.
  • Saved and reloaded model weights for inference, qualitatively and quantitatively comparing denoised outputs against original images to assess reconstruction quality.

Autoencoder Denoising PyTorch CNN Conv2d ConvTranspose2d CIFAR-10 Image Reconstruction Representation Learning Computer Vision Deep Learning Model Serialisation

Digital Forensics Investigation Lab

  • Analysed disk images and memory dumps using Autopsy, FTK, and Volatility to recover digital artefacts and reconstruct event timelines in Windows environments.
  • Captured and inspected network traffic with Wireshark to identify indicators of compromise, complemented by password analysis using John the Ripper.
  • Documented findings following structured forensic procedures, maintaining chain of custody and producing professional evidence reports suitable for investigative use.

Digital Forensics Memory Forensics Disk Forensics Autopsy FTK Volatility Wireshark John the Ripper Network Traffic Analysis Incident Response Chain of Custody Cybersecurity

SoK: Intrusion Detection for Automotive and Autonomous Vehicle Systems

  • Co-authored a Systematization of Knowledge (SoK) paper surveying Intrusion Detection Systems for automotive networks, covering CAN bus and V2X communication protocols and their security vulnerabilities.
  • Developed a two-dimensional IDS taxonomy categorising 20+ systems by algorithm type (timing, rule, entropy, specification, ML, and deep learning) and attack type under the CIA triad, analysing systems including CANShield, EdgeTDC, and MULSAM.
  • Critically evaluated trade-offs between detection accuracy, computational cost, and real-world deployability, identifying open challenges and proposing future directions including federated learning and security-by-design for next-generation protocols.

Cybersecurity Intrusion Detection Automotive Security CAN Bus V2X Autonomous Vehicles Machine Learning Deep Learning Federated Learning Taxonomy Research SoK Academic Writing CIA Triad

BSc Year 4 Projects

A Study on Buffer Overflow Attacks

  • Exploited stack-based buffer overflows in vulnerable C programs across Windows XP, Windows 11, and Linux to achieve arbitrary code execution, demonstrating cross-platform offensive security fundamentals.
  • Wrote Bash scripts on Kali Linux to open a remote shell on a Windows 11 target, simulating a real-world post-exploitation scenario.
  • Used GDB and pwntools to debug binaries and develop proof-of-concept exploits, evaluating mitigations including stack canaries, ASLR, and DEP/NX.

Buffer Overflow Exploit Development Offensive Security GDB pwntools Kali Linux C ASLR Penetration Testing

ANN with Particle Swarm Optimisation

  • Implemented an Artificial Neural Network and Particle Swarm Optimisation algorithm entirely from scratch in Python, without any external ML libraries.
  • Designed an interactive Tkinter GUI to visualise training progress and model behaviour in real time, enabling intuitive exploration of the optimisation process.
  • Structured the project into modular components, cleanly separating optimisation logic, model architecture, and the user interface for maintainability and clarity.

Artificial Neural Networks Particle Swarm Optimisation Python Tkinter

BSc Year 3 Projects

MemoriesAR — Full-Stack Mobile Application

  • Co-developed a full-stack platform with a mobile app and web admin panel using React, React Native, Node.js, MySQL, and Firebase Authentication.
  • Built and tested backend APIs for authentication and database access, ensuring reliable integration between frontend clients and server-side services.
  • Worked in an Agile/Scrum environment, contributing to sprint planning and iterative feature development across a team.

React React Native Node.js MySQL Firebase REST APIs Full-Stack Development Mobile Development Agile Scrum Team Collaboration

BSc Year 2 Projects

Mastermind Game — C & ARM Assembly on Raspberry Pi

  • Implemented a Mastermind game in C with inline ARM Assembly on a Raspberry Pi, interfacing with physical components including LEDs, LCD display, and buttons via GPIO.
  • Managed low-level register usage, execution flow, and direct hardware interaction at the assembly level, maintaining correctness across C and assembly boundaries.
  • Debugged complex cross-language behaviour in a hardware-software co-design context, demonstrating hands-on embedded systems development with real physical components.

ARM Assembly C Raspberry Pi Embedded Systems Hardware Programming Low-Level Programming GPIO LCD Display Inline Assembly Hardware-Software Interface

LSB Steganography Tool in C

  • Built a command-line steganography tool in C to encode and decode hidden messages within PPM images using Least Significant Bit (LSB) encoding.
  • Implemented bit manipulation logic to embed secret data into image pixel values with minimal visual distortion, preserving image integrity across encode-decode cycles.
  • Developed robust error handling to manage edge cases such as oversized messages, malformed image files, and invalid inputs, ensuring reliable CLI operation.

[Code]

C Steganography LSB Encoding PPM Images CLI Image Processing Low-Level Programming Bit Manipulation Error Handling