Search Results (2)

Through the addition of humanly imperceptible noise to an image classified as belonging to a category $c_a$, targeted adversarial attacks can lead convolutional neural networks (CNNs) to classify a modified image as belonging to any predefined target class $c_t\ne c_a$. To achieve a better understanding of the inner workings of adversarial attacks, this study analyzes the adversarial images created by two completely opposite attacks against 10 ImageNet-trained CNNs. A total of $2\times 437$ adversarial images are created by ${\mathrm{EA}}^{\mathrm{target},\mathcal{C}}$, a black-box evolutionary algorithm (EA), and by the basic iterative method (BIM), a white-box, gradient-based attack. We inspect and compare these two sets of adversarial images from different perspectives: the behavior of CNNs at smaller image regions, the image noise frequency, the adversarial image transferability, the image texture change, and penultimate CNN layer activations. We find that texture change is a side effect rather than a means for the attacks and that $c_t$-relevant features only build up significantly from image regions of size $56\times 56$ onwards. In the penultimate CNN layers, both attacks increase the activation of units that are positively related to $c_t$ and units that are negatively related to $c_a$. In contrast to ${\mathrm{EA}}^{\mathrm{target},\mathcal{C}}$’s white noise nature, BIM predominantly introduces low-frequency noise. BIM affects the original $c_a$ features more than ${\mathrm{EA}}^{\mathrm{target},\mathcal{C}}$, thus producing slightly more transferable adversarial images. However, the transferability with both attacks is low, since the attacks’ $c_t$-related information is specific to the output layers of the targeted CNN. We find that the adversarial images are actually more transferable at regions with sizes of $56\times 56$ than at full scale. Full article

The elementary processes of anionic styrene polymerization in the gas phase and in cyclohexane were studied using M062X (a recently developed density functional theory (DFT) method) combined with the 6-31+G(d) basis sets, in order to clarify the complicated phenomena caused by the association of the active chain-ends and elucidate the details of the polymerization mechanism. Three types of HSt₂Li (a model structure of polystyryllithium chain-ends) were obtained; the well-known first structure in which Li is coordinated to the side chain, the second structure in which Li is coordinated to the phenyl ring, (both without the penultimate unit coordination), and the third structure in which Li is coordinated to both the chain-end unit and the penultimate styrene unit. Although the third HSt₂Li is the most stable as expected, the free energy for the transition state of its reaction with styrene is higher than those for the other two transition states due to its steric hindrance. The free energy for the transition state of the reaction of the second HSt₂Li with styrene is the lowest, suggesting that the route through it is the predominant reaction path. The penultimate unit effect, slower addition of styrene to HSt₂Li than to HStLi, is attributed to coordination of the penultimate styrene units of the polystyryllithium dimer (one of the starting materials) to its Li atoms. The calculated enthalpy for the reaction barrier of the second HSt₂Li with styrene in cyclohexane was found to agree with the observed apparent activation energy in benzene. Full article