Generation of Environment-Irrelevant Adversarial Digital Camouflage Patterns.

Digital camouflage is the most common and effective means to combat military reconnaissance. Traditional digital camouflage generation methods must regenerate camouflage images according to the current environment. When the environment changes, generated camouflage images may be detected by neural network classification models. We present a digital camouflage generation model based on disentangled representation, which can decompose images into a content space and a style space, thereby recombining the current content of the environment image with different digital camouflage styles. When the environment changes, our model can generate digital camouflage images based on the original environment content and the corresponding digital camouflage style, without obtaining the current environment image. To counter the detection of the classification models, we design a category reordering function to mislead the classification result of the classification model. Experiments show that the proposed method can generate digital camouflage images in different seasons and successfully implement an adversarial attack on the classification model.