Karthikeyan Saravanan

The paper proposes PINA, a two-stage differentially private clustered federated learning framework that improves convergence and robustness by using low-rank adaptation and a normality-driven aggregation mechanism.

The paper proposes DP-LAC, a novel lightweight adaptive clipping technique for differentially private federated fine-tuning, which efficiently estimates and adapts the clipping threshold without consuming extra privacy budget or requiring manual hyperparameter tuning.

DisAgg introduces a novel secure aggregation protocol that uses a small committee of Aggregators to compute partial sums, achieving a significant speedup (4.6x) over previous state-of-the-art methods like OPA while maintaining privacy.