Omar Hamdash

Omar Hamdash

Projects

🕜 in progress

Blockchain Healtcare Data Management System

By leveraging the inherent advantages of blockchain technology, this system aims to provide a secure, transparent, and efficient method for managing patient data while ensuring privacy and controlled access.

Details

Each patient will be able to upload their medical data to the system of distributed ledger and allow access to other parties based on their needs. By utilizing smart contracts it is possible to have high auditability and accountability.

Main features:

Interoperability, transparency, encrypted medical data, auditablity, Role-Based Access Control (RBAC), grant or revoke access to data, decentralized data management.

Readings

General info - survey - interoperability/security

🕜 in progress

MediXAI

MediXAI is an end-to-end doctor and patient assistant for early breast cancer screening. Empowered by XAI (explainable AI), it helps doctors, especially radiologists, to diagnose breast cancer at early stages with higher accuracy. The flexibility and accessibility that MediXAI provides make it easier for patients to follow their screening results and contact with their doctors. The availability of many doctors on one platform allows patients to get access to many opinions and therefore, satisfying the desire of the second opinion market.

Machine Learning:

We trained our machine learning model to classify mammograms into the 5 BI-RADS classes. The model also offers explainable results where the doctors can check the areas where tumor is found and therefore eliminating the blackbox issue where doctors don't understand what the classification was based on.

Data Privacy:

Methods used: RSA 2048 bits, AES-256, hashing, PBKDF2.
Since the data handled is sensitive medical data, we implemented an ecnryption scheme that allows for the collaboration of doctors and the sharing of medical data. The encryption scheme is end-to-end and client side, preventing even the service providers to access sensitive data.

Technologies used:

Resnet, VGG, GRAD-CAM, AWS, Flutter, Flask, Cryptography

Technologies to be implemented:

Inference on encrypted mammograms, blockchain data management system with off-chain storage model, the ability to grant and revoke access to resources.

Links:

🌐Project's website - Doctor's web app - Patient's Android apk

🕜 in progress

Vector Database as a Service

The vector database offers high-performance, scalable database optimized for storing, indexing, and querying high-dimensional vector data. This database is particularly suited for applications involving machine learning, artificial intelligence, and data science, where handling large volumes of vector data efficiently is crucial.

Key features:

Storing and queriying vector data, storage of metadata related to the vector data, indexing of vector data using HNSW (Hierarchical Navigable Small World Graph).

Main use cases:

Vector databases are particularly useful in the process of fine tuning Large Language Models (LLM) with specific data. Vector embeddings are stored in the vector DB, then a method called Retrieval-Augmented Generation (RAG) is used to generate responses based on the previously stored data.

ML Natural Adversarial Examples Defense

Natural adversarial examples present a significant challenge to the robustness of deep neural networks (DNNs) in real-world ap- plications. These examples arise from natural variations within datasets and are not artificially generated. This paper introduces a novel defense approach against natural adversarial examples in the ImageNet dataset by leveraging Salient Feature Extraction (SFE). Our method distinguishes between salient features (SF), which are robust and aligned with human perception, and trivial features (TF), which often mislead models. Utilizing a coupled generative adversarial network (GAN), we effectively extract and prioritize SFs, thereby enhancing the model’s ability to accurately classify and defend against natural adversarial examples

Links:

Report

Video stabilization using face tracking

In this project, we aim to obtain a stabilized video by eliminating unwanted camera motions. For that purpose, instead of just suppressing high-frequency motion paths, we employed the method mentioned in the article Auto-Directed Video Stabilization with Robust L1 Optimal Camera Paths. We also added face tracking to further improve the proposed metod and make the stabilized video smoother in applications where faces are present in the video to stbilize. The stabilization process happens in three main steps: firstly estimating the original shaky camera path, secondly estimating a new smooth camera path, and thirdly reproducing the stabilized video using the estimated smooth camera path.

Links:

Report - Github

Distributed File System

A distributed file system that runs on the local area network. Being scalable, it allows as many servers and clients as desired to use the system at once. A master server functions as a middle-ware. The system uses ZMQ for messaging, allowing for distributed communication between different nodes in the system

Technologies Used

ZMQ, Networking, LAN networking, sockets, Linux VMs

Links:

GitHub

GoBooking

A user-friendly property booking application designed to streamline the process of searching, booking, and reviewing accommodations. The platform offers a range of features and functionalities to cater to the diverse needs of users, homeowners, and administrators.

Technologies Used

Spring Boot, React, PostgreSQL

Links:

GitHub

BilkE

BilkE is an ERASMUS application and management system. The BilkE api manages a database and uses Amazon AWS S3 services to store files on the cloud. Other operations are done by the api like email verification and token generation and management. The system allows Bilkent University students to manage their ERASMUS exchange applications. ERASMUS coordinators can also do all the jobs they used to do manually on the system. The main goal from this system was to digitize the process of ERASMUS applications from Bilkent University and eliminate all non digital documents and forms.

Technologies Used

AWS S3, Spring Boot, React, PostgreSQL

Links:

GitHub