Research Article

AEGuard: Image Feature-Based Independent Adversarial Example Detection Model

Table 2

Proposed system training dataset attributes.

AttributeDatatypeCorrelationExplanation

edge_noiseDouble0.6974Ratio of noise to the edge pixels
edge_densityDouble0.5450Ratio of edge pixels to the total pixels of image
dct_biasDoubleāˆ’0.8172DCT bias of the image
color_compositionDouble (3)0.1193Percentage of each color channel (R, G, and B) of the image
VarianceDouble0.2620Image variance of the image
EntropyDouble0.1522Image entropy of the image
is_advBooleanā€‰Normal image/adversarial example