Errata

IEEE SIgnal ProcESSIng MagazInE | January 2018 | I n the article “The Robustness of Deep Networks” [1] by A. Fawzi, S.-M. Moosavi-Dezfooli, and P. Frossard in the November 2017 issue of IEEE Signal Processing Magazine, several references were cited incorrectly and [48] was inadvertently omitted from the “References” section. On page 60 of the article, the correct references are cited as follows: ■ Nayebi and Ganguli consider a regularizer to push activations of the network in the saturating regime of the nonlinearity (i.e., the region where the nonlinear activation function is flat) [46]. ■ In [47], a regularization scheme is introduced for improving the network’s sensitivity to perturbations by constraining the Lipschitz constant of the network. In [41], an informationtheoretic loss function is used to train stochastic neural networks; the resulting classifiers are shown to be more robust to adversarial perturbations than their deterministic counterpart. On page 61, the correct reference is: ■ To build robust deep representations, [48] considers instead a new architecture with fixed filter weights. On page 62 of the article, [48] was excluded. It is as follows: ■ [48] J. Bruna and S. Mallat, “Invariant scattering convolution networks,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 8, pp. 1872–1886, 2013. We regret these errors and apologize for any confusion they may have caused.