Elina Israyelyan, Anri Abrahamyan, Hovhannes Manushyan

Disease Classification on Encrypted Retina Scan Images using Neural Network


Elina Israyelyan is a Software Engineer in Picsart. She has more than 2 years of experience in machine learning and engineering specializing in Computer vision and NLP.

Anri Abrahamyan is a DevOps engineer in Align Technology. Besides pursuing his career in Ops he has 2 years of experience in machine learning and AI research. 

Hovhannes Manushyan is a Data Scientist in Disqo. He has a more than 2 years of experience in Data science specializing in recommender systems.


Description of the Talk:

The capstone project utilizes homomorphic encryption to ensure privacy while detecting diseases on retinal scan images. They employ convolutional neural networks and vision transformer models, demonstrating the effectiveness of homomorphic encryption and addressing encryption-related challenges. Their experiments yield promising results in accuracy and privacy preservation, with their vision transformer achieving 83% accuracy on the plain testing dataset. They have also developed an approximation function for layer normalization, bringing us closer to performing fully encrypted inference with vision transformers.