At American University of Beirut, I built reproducible machine-learning training and evaluation pipelines using PyTorch, Opacus, and DP-SGD. My research focused on privacy-preserving training and model performance across groups.
I designed a private group-progress controller that used clipped, noise-perturbed group-loss signals to guide adaptive learning rates while preserving privacy accounting. On skewed MNIST, this raised digit-8 accuracy from 60.40% to 65.30% and narrowed the majority-minority accuracy gap under the same privacy budget.
For a controlled Adult dataset experiment, adaptive differentially private optimization improved overall accuracy from 76.30% to 83.93% and balanced accuracy from 54.08% to 77.10% versus a fixed-learning-rate DP baseline. My work also includes a published UAV GPS spoofing detection project, where I used genetic-algorithm hyperparameter search for an LSTM classifier.
I also worked on a football outcome and score prediction project, building a leakage-aware match-outcome pipeline and owning the statistical-modelling track. I hold a Master of Engineering in Electrical and Computer Engineering from American University of Beirut and am certified in NVIDIA Fundamentals of Deep Learning.
