notesum.ai

Published at November 11

Computable Model-Independent Bounds for Adversarial Quantum Machine Learning

cs.LG

cs.AI

cs.ET

quant-ph

Released Date: November 11, 2024

Authors: Bacui Li¹, Tansu Alpcan¹, Chandra Thapa², Udaya Parampalli³

Aff.: ¹Department of Electrical and Electronic Engineering, University of Melbourne; ²CSIRO Data61; ³School of Computing and Information Systems, University of Melbourne

Arxiv: http://arxiv.org/abs/2411.06863v1

MNIST models
Model	Non-Adv.	Attack	Adv.	Estimated	Error $>$ Bound
Model	error	Str. ( $\epsilon$ )	error	Bound	Error $>$ Bound
M1	0.2543	0.1 ( $l_{q}$ )	0.617	0.307	True
M1	0.2543	100 ( $l_{2}$ )	0.376	0.273	True
M2	0.1601	0.1 ( $l_{q}$ )	0.501	0.186	True
M2	0.1601	100 ( $l_{2}$ )	0.243	0.168	True
M3	0.0772	0.1 ( $l_{q}$ )	0.564	0.084	True
M3	0.0772	100 ( $l_{2}$ )	0.140	0.082	True
FashionMNIST models
Model	Non-Adv.	Attack	Adv.	Estimated	Error $>$ Bound
Model	error	Str. ( $\epsilon$ )	error	Bound	Error $>$ Bound
F1	0.4207	0.1 ( $l_{q}$ )	0.655	0.499	True
F1	0.4207	100 ( $l_{2}$ )	0.499	0.442	True
F2	0.3179	0.1 ( $l_{q}$ )	0.526	0.358	True
F2	0.3179	100 ( $l_{2}$ )	0.362	0.331	True
F3	0.2228	0.1 ( $l_{q}$ )	0.673	0.260	True
F3	0.2228	100 ( $l_{2}$ )	0.306	0.237	True