Research Article
Improving the Imperceptibility of Adversarial Examples Based on Weakly Perceptual Perturbation in Key Regions
| Initialization: T = Faster RCNN, f = Rf, dataset = VOC 2007, batchsize = 1, epoch = 5, α = 1000, β = [0.0001, 0.0002], and γ = 1 | | While t < epoch: | | for x in the dataset: | | Train G: | | G (x) ⟶ r, M (x)⟶M, | | F (p) ⟶ P | | x′ = x + P | | T (x′) ⟶ | | G (x, p) ⟶ LGAN_G | | Calculate Lperceptual (x, x′), Lfeature (f,), and LDAG | | LG = LGAN_G + αLperceptual + βLfeature + γLDAG | | Update G parameters | | Train D: | | D (img_real, ones) ⟶ loss_real | | D (img_fake, zeros) ⟶ loss_ fake | | LD = (loss_real + loss_ fake)0.5 | | Update D parameters | | end for | | t ⟶ t++ | | end while |
|