XAI 기반 생성 모델 데이터 분석 및 평가

Abstract

When performing guided pilot simulations or flight test simulations required for var- ious purposes, the quality of dynamic and synthetic sensor images generated from the sensor models in the given simulation environment is highly important for target recog- nition, tracking, and behavior for various reconnaissance missions. So this thesis uses artificial intelligence to generate realistic infra-red (IR) images used in flight simulations, and analyzes and evaluates the data through explainable artificial intelligence (XAI). We construct an IR dataset and a synthetic dataset to generate realistic IR images. Among the Generative Adversarial Network (GAN) models, CycleGAN, which is trained under unpaired dataset, is trained using the constructed dataset. CycleGAN produces high-quality images even though it does not have a correct answer label. However, it is not well-trained on the IR dataset. Therefore, we improved the model’s performance based on the structural similarity index measure (SSIM). At this time, we compared the weights of each loss function to find an appropriate value, and analyzed how window sizes of SSIM would affect the synthetic IR image constructed by CycleGAN is analyzed. Although techniques such as IS and FID have been introduced to evaluate the per- formance of GAN, it is still difficult to distinguish between synthetic data. Additionally, distinguishing synthetic data generated by artificial intelligence is a big topic because the level of data generation using GAN is improving. Therefore, we introduce XAI for synthetic IR image analysis. Out of various XAI techniques, LRP was used, which detects the model in reverse order through decomposition and relevance propagation, providing a basis for judgment on prediction. Thus, we build a classification network for IR images and synthetic IR images and then perform LRP analysis. When analyz- ing LRP, we simulate various transformations of LRP and analyze how LRP draws a heatmap when some transformation is applied to the data, allowing us to distinguish between IR images and synthetic IR images.