Publications

FIA, a training-free framework for multi-concept unlearning in diffusion models leveraging model sparsity. Introduces Contrastive Concept Saliency to quantify each weight's contribution to a target concept, identifies Concept-Sensitive Neurons through temporal and spatial analysis, and fuses per-concept masks into a unified multi-concept mask. Operates at under 0.3% overall sparsity and achieves SOTA across multi-object, NSFW, and artistic-style unlearning tasks.

ScaPre, a unified closed-form framework for scalable and precise large-scale concept unlearning. Combines a conflict-aware stable design (spectral trace regularization + geometry alignment) with an Informax Decoupler that confines updates to concept-relevant subspaces. Without extra data or auxiliary modules, ScaPre unlearns up to 5× more concepts than the best baseline, completing 50 concepts in 120 seconds.

A novel unlearning framework combining a Dynamic Mask and a Concept-Aware Loss for multi-concept forgetting in text-to-image diffusion models. The dynamic mask adaptively updates gradient masks during optimization for selective weight modifications; the concept-aware loss enforces semantic consistency via superclass alignment and uses distillation to prevent unlearned concepts from resurfacing.

GS2E, a large-scale synthetic event dataset for high-fidelity 3D event-based vision, comprising 1,150+ scenes. Reconstructs photorealistic static scenes via 3D Gaussian Splatting from sparse multi-view RGB images, then runs a physically-informed event simulation pipeline with adaptive trajectory interpolation and contrast-threshold modeling.

SecYOLOv7, a secure single-stage detector for privacy-preserving transmission-line inspection. Uses Shamir secret-sharing-based MPC across two non-colluding edge servers to perform Secure Feature Contraction, Secure Bounding-Box Prediction, and Secure Object Classification. Maintains computation error around 1e-4 with 2.113s runtime and 95.15 MB communication overhead.