2025 - PresentOngoing

Adaptive and Hybrid Privacy-Preserving Framework for Text-Based Machine Learning to Secure Training Data

University of Colombo School of Computing

Building a modular hybrid PPML platform for text classification that can switch between FL, DP, HE, and SMPC at runtime. The goal is to defend against adaptive privacy attacks (e.g., membership inference, model inversion, reconstruction) by dynamically tuning privacy/security parameters while keeping model utility and system efficiency practical for real deployments.

• Runtime Adaptation Architecture to plug-and-play FL / DP / HE / SMPC and switch methods based on detected threats
• RL-driven parameter adaptation (e.g., DP epsilon, encryption/security parameters, aggregation frequency) using system metrics as state and privacy–utility as reward
• Attack-aware evaluation using privacy/utility + semantic preservation metrics (AOP, SFS), computational overhead, and reduction in attack success rates

2025 - PresentOngoing

Explainable ML Forecasting for Monthly USD/LKR Exchange Rate

Developing an accurate and transparent USD/LKR monthly forecasting framework by combining modern machine-learning models with SHAP explainability. The study builds a unified macroeconomic dataset (domestic + global), compares statistical and ML models (ARIMA, XGBoost, LSTM, hybrid), and extracts policy-relevant insights especially around crisis regimes such as Sri Lanka’s 2022 volatility.

• Unified monthly dataset with key macroeconomic drivers (inflation, interest rates, FX reserves, oil prices, US rates, etc.) aligned for forecasting
• Comparative modeling: ARIMA baseline vs multivariate ML (XGBoost, LSTM) plus a hybrid/ensemble approach
• Rolling-window / out-of-sample evaluation using RMSE, MAE, MAPE, and directional accuracy with significance checks when applicable