Publications

We propose DPZV, the first zeroth-order (ZO) optimization framework for vertical federated learning (VFL) that achieves tunable differential privacy (DP) with formal guarantees and low memory overhead by injecting low-variance scalar noise at the server, enabling privacy-preserving collaborative training without explicit gradient exposure, matching the convergence rate of first-order methods, and outperforming existing baselines on image and language benchmarks across a wide range of privacy constraints.

We propose MU-SplitFed, a straggler-resilient Split Federated Learning (SFL) algorithm that addresses the critical bottleneck of client-server synchronization delays—caused by SFL’s dependency on client-side activations—by introducing a simple yet effective unbalanced update mechanism that enables per-client local updates on the server, achieving a linear reduction in communication rounds and demonstrating superior performance over baselines under straggler conditions through adaptive tuning.

We investigate Federated Foundation Models to bring reliable, adaptive intelligence to harsh environments-where connectivity is intermittent, devices are resource-limited, and data is noisy—by combining FM generalization with decentralized, communication-aware FL. Our work maps key system constraints and outlines research challenges in communication design, robustness, and energy-efficient personalization.

We introduce Parameter Tracking (PT), a generalization of Gradient Tracking (GT) for Federated Learning (FL), and propose two novel Adam-based FL algorithms, FAdamET and FAdamGT, that integrate PT for improved adaptive optimization. Theoretical analysis and experiments demonstrate that these methods enhance convergence efficiency, reducing both communication and computation costs across varying data heterogeneity levels.

A privacy-enhanced FL framework that optimally injects Differential Privacy noise at different hierarchical layers of fog nodes on untrusted models within subnetworks. Our method achieves strong privacy guarantees while maintaining superior model performance and convergence efficiency compared to baseline methods.

A DP-secured FL framework that adapts noise injection across multiple tiers of networks based on heterogeneous trust models. Our method optimizes privacy and performance, achieving significant improvements in convergence, energy, latency, and model quality across various privacy budgets and system configurations.

Hierarchical Federated Learning (HFL) faces challenges from multi-timescale model drift due to data heterogeneity across hierarchical levels. To address this, we propose Multi-Timescale Gradient Correction (MTGC), achieving stable convergence independent of data heterogeneity and demonstrating superior performance in diverse HFL settings..

We propose Semi-Decentralized Gradient Tracking (SD-GT) to enhance federated learning over fog networks by eliminating gradient diversity assumptions, enabling efficient hierarchical model aggregation and achieving superior model quality with reduced communication costs.

The first Semi-Decentralized Federated Learning (SD-FL) framework that eliminates the need for data heterogeneity assumptions by incorporating tracking terms into device updates. Our method achieves significant improvements in model quality and communication efficiency over existing SD-FL methods.

A framework that combines super-resolution techniques with fovea rendering to enable efficient low-bandwidth video streaming over low-power protocols like Bluetooth. By leveraging deformable convolutional networks (DCN), our model achieves state-of-the-art video quality with fast, low-energy performance, making it ideal for VR/AR applications.

A meta-learning approach for remote heart rate estimation using rPPG, enabling self-supervised weight adjustment during deployment to adapt to distributional shifts. This method improves model robustness to variations in skin tone, lighting, and facial structure.