notesum.ai

Published at November 21

Indiscriminate Disruption of Conditional Inference on Multivariate Gaussians

stat.ML

cs.CR

stat.AP

Released Date: November 21, 2024

Authors: William N. Caballero¹, Matthew LaRosa², Alexander Fisher³, Vahid Tarokh²

Aff.: ¹Department of Operational Sciences, Air Force Institute of Technology; ²Department of Electrical and Computer Engineering, Duke University; ³Department of Statistical Science, Duke University

Arxiv: http://arxiv.org/abs/2411.14351v1

^‡ WB, RN and $\mathbf{z}^{\prime}$ values correspond to whitebox objective; SAA and SGA to grey-box
Attack	$\left(\boldsymbol{\mu}_{0},\kappa,\boldsymbol{\Psi}\ ,\nu\right)$	$z_{1}$	$z_{2}$	$z_{3}$	$z_{4}$	$z_{5}$	$z_{6}$	$z_{7}$	$z_{8}$	$z_{9}$	$z_{10}$	$z_{11}$	Obj. Value^‡	$D_{KL}$ ^∗∗
SAA	$\left(\hat{\boldsymbol{\mu}},5,2\hat{\boldsymbol{\Sigma}}+\mathbf{I},18\right)$	1.609	1.301	1.194	1.527	1.092	1.681	0.989	0.917	0.829	1.007	0.604	0.0013	58.377
	$\left(\hat{\boldsymbol{\mu}},10,3\hat{\boldsymbol{\Sigma}},19\right)$	1.609	1.065	1.194	1.527	0.858	1.681	0.989	0.917	0.801	0.707	0.506	-0.0429	257.574
	$\left(0.5\hat{\boldsymbol{\mu}},1,0.5\hat{\boldsymbol{\Sigma}},18\right)$	1.609	1.065	1.194	1.246	0.841	1.681	0.989	0.917	0.822	0.707	0.388	-0.9962	309.905
	$\left(2\hat{\boldsymbol{\mu}},10,2\hat{\boldsymbol{\Sigma}},36\right)$	1.417	1.365	1.194	1.527	0.792	1.681	0.938	0.877	0.829	0.707	0.688	1.457	38.4578
\cdashline1-15 RN	-	1.494	1.073	1.241	1.431	0.920	1.580	0.959	0.880	0.738	0.798	0.427	-0.233	108.998
\cdashline1-15 $\mathbf{z}^{\prime}$	-	1.459	1.214	1.344	1.377	0.941	1.531	0.838	0.766	0.679	0.856	0.537	-0.427	0
^∗∗ KL divergence between induced and true conditional treating the white-box model as the true joint