notesum.ai

Published at November 2

Privacy-Preserving Federated Learning with Differentially Private Hyperdimensional Computing

cs.LG

cs.AI

cs.CR

stat.ML

Released Date: November 2, 2024

Authors: Fardin Jalil Piran¹, Zhiling Chen¹, Mohsen Imani², Farhad Imani¹

Aff.: ¹School of Mechanical, Aerospace, and Manufacturing Engineering, University of Connecticut; ²Department of Computer Science, University of California, Irvine

Arxiv: http://arxiv.org/abs/2411.01140v1

Notation	Description
$\mathbf{X}^{s}_{k,r}$	Training data generated by client $k$ for class $s$ at round $r$
$\vec{H}^{s}_{k,r}$	Hypervectors from client $k$ for class $s$ at round $r$
$\vec{C}^{s}_{k,r}$	Class $s$ hypervector for client $k$ at round $r$
$\{{\vec{C}}^{s}_{g,r}\}_{s=1}^{S}$	Global model at round $r$
$\{\tilde{{\vec{C}}}^{s}_{k,r}\}_{s=1}^{S}$	Secure local model for client $k$ at round $r$
$\{\tilde{{\vec{C}}}^{s}_{g,r}\}_{s=1}^{S}$	Secure global model at round $r$
$\sigma_{dp}^{2}$	Noise level (Variance)
$\Delta$	Sensitivity
$\epsilon$	Privacy budget
$\delta$	Privacy loss threshold
$\vec{\xi}_{k}^{r}$	Required noise for client $k$ ’s local model at round $r$
$\vec{\Psi}_{k}^{r}$	Cumulative noise in client $k$ ’s local model at round $r$
$\vec{\Gamma}_{k}^{r}$	Additional noise added to client $k$ ’s local model at round $r$
$\vec{\xi}_{g}^{r}$	Required noise for the global model at round $r$
$\vec{\Psi}_{g}^{r}$	Cumulative noise in the global model at round $r$
$\vec{\Gamma}_{g}^{r}$	Additional noise added to the global model at round $r$