Projects

A resource-adaptive federated fine-tuning method that fixes the expert-imbalance and gradient-sparsity discordances of sparse MoE, achieving up to 45% computation reduction and 8.7x better low-resource performance over heterogeneous LoRA-rank methods.

A semi-supervised detector that learns architectural patterns in fake images and generalizes across image generators.

A survey of disruption and watermarking methods for proactive deepfake defense across visual and audio media.

A learning-to-unlearn framework that incorporates transformations during federated learning to reduce memorization and simplify subsequent unlearning.

A test-time adaptation method that helps deepfake detectors adjust during inference without training data or labels.

A cancellable biometric hashing scheme for voice-based speaker verification that records element rankings instead of maximum values.

We test D-CAPTCHA against transferable, imperceptible adversarial attacks and strengthen it through adversarial training.

A framework that combines differential privacy and meta-learning to address privacy leakage and non-IID data in cross-silo federated learning.