LATIDIA · Ciberseguridad

Recorte sensible a la curvatura: mitigación del sesgo en DPSGD a través de la información aproximada de segundo pedido

arXiv:2610.32175v1 Announce Type: cross Abstract: Differentially Private Stochastic Gradient Descent (DPSGD) se ha convertido en el estándar de facto para entrenar redes neuronales profundas con rigurosas garantías de privacidad. Sin embargo, t

La noticia

arXiv:2610.32175v1 Announce Type: cross Abstract: Differentially Private Stochastic Gradient Descent (DPSGD) has become the de facto standard for training deep neural networks with rigorous privacy guarantees. However, the utility of DPSGD is constrained by gradient clipping bias and DP noise. While existing work has tried to address these issues, the geometric misalignment caused by clipping bias remains a critical challenge. To bridge this gap, we propose a novel framework, Curvature-Aware Differentially Private Stochastic Gradient Descent (CA-DPSGD), which leverages approximate second-order information to construct an optimal, curvature-aware clipping space. CA-DPSGD integrates a rectified estimator for the curvature matrix and is coupled with an anisotropic noise mechanism to ensure privacy guarantees. Extensive