Jason Li

Publication Date

Fall 2025

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Fabio Di Troia

Second Advisor

Amith Kamath Belman

Third Advisor

Sayma Akther

Keywords

Facial Recognition, Machine Learning, Diffusion, InsightFace, Gen- erative AI

Abstract

In the realm of facial recognition biometric systems, there are many challenges correlated with the robustness of the user templates in the system. Low sample counts in user image templates can weaken biometric recognition systems, reducing accuracy and robustness. Using generative AI, these challenges can be mitigated through the creation of realistic synthetic user images to introduce additional samples to user templates. Previous research focused on using generative adversarial networks (GAN) and variational autoencoders (VAE) as their generative models of choice. This work utilizes a conditional diffusion model, a newer technology in the realm of generative AI. The model was trained on the PubFig83 dataset, a dataset of 13838 cropped images of 83 public figures. The model the generated images by feeding in noised images from the dataset alongside an identity embedding to guide the model to generate images containing features of specified figures. A pre-trained facial feature extractor and embedding generator model pack from InsightFace was utilized. Cosine similarity alongside machine learning approaches like SVCs and KNNs utilized these embeddings for training and testing. These approaches were provided with various datasets of increasing percentages of synthetic recreated images, ranging from 0 to 100% in increments of 10%. Additional low sample experiments were also explored. The primary objective was to see if diffusion models could increase user template robustness through data augmentation. The results show promise that diffusion model synthetic recreations contain the necessary identity features for facial recognition.

Download

Available for download on Saturday, December 19, 2026

Share

COinS