<ul><li>The paper proposes a framework called PCDP-SGD to improve the convergence of Differentially Private Stochastic Gradient Descent (DP-SGD).</li><li>PCDP-SGD compresses redundant gradient norms and preserves crucial top gradient components through a projection operation before gradient clipping.</li><li>The framework is extended to differential privacy federated learning (DPFL) to tackle data heterogeneity and achieve efficient communication.</li><li>Experimental results show that PCDP-SGD achieves higher accuracy compared to other DP-SGD variants and outperforms current federated learning frameworks in computer vision tasks.</li></ul>

PCDP-SGD: Improving the Convergence of Differentially Private SGD via Projection in Advance

Discover more