• IEMEK Journal of Embedded Systems and Applications
• /
• v.16 no.4
• /
• pp.145-151
• /
• 2021

In this paper, we propose a novel visualization technique to explain the predictions of deep neural networks. We use knowledge distillation (KD) to identify the interior of a black-box model for which we know only inputs and outputs. The information of the black box model will be transferred to a white box model that we aim to create through the KD. The white box model will learn the representation of the black-box model. Second, the white-box model generates attention maps for each of its layers using Grad-CAM. Then we combine the attention maps of different layers using the pixel-wise summation to generate a final saliency map that contains information from all layers of the model. The experiments show that the proposed technique found important layers and explained which part of the input is important. Saliency maps generated by the proposed technique performed better than those of Grad-CAM in deletion game.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2014.06a
• /
• pp.105-106
• /
• 2014

본 논문에서는 중요도 지도를 사용한 화소값 사이 척력 기반 영상 대조비 향상 기법을 제안한다. 공간상에서 인접한 화소들 사이에 자주 발생하는 화소값들의 차이를 크게 하면 효과적으로 영상의 디테일을 두드러지게 할 수 있다. 대조비 증가를 위해 화소값 사이 척력을 정의하고, 유효 화소값 사이 척력들의 합을 사용하여 대조비의 증가 정도를 조절한다. 중요도 지도는 영상의 화소마다 사람의 시선이 머무르는 정도를 상대적인 수치로 나타낸 것이다. 따라서 영상 화질을 개선할 때 중요도 지도를 사용하면 사람의 시선을 끄는 화소값들의 대조비를 선택적으로 높일 수 있다. 실험 결과를 통하여 제안 기법이 우수한 화질개선 영상을 생성함을 확인한다.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.12 no.6
• /
• pp.629-635
• /
• 2019

In Deep Learning models derivative is implemented by error back-propagation which enables the model to learn the error and update parameters. It can find the global (or local) optimal points of parameters even in the complex models taking advantage of a huge improvement in computing power. However, deliberately generated data points can 'fool' models and degrade the performance such as prediction accuracy. Not only these adversarial examples reduce the performance but also these examples are not easily detectable with human's eyes. In this work, we propose the method to detect adversarial datasets with random noise addition. We exploit the fact that when random noise is added, prediction accuracy of non-adversarial dataset remains almost unchanged, but that of adversarial dataset changes. We set attack methods (FGSM, Saliency Map) and noise level (0-19 with max pixel value 255) as independent variables and difference of prediction accuracy when noise was added as dependent variable in a simulation experiment. We have succeeded in extracting the threshold that separates non-adversarial and adversarial dataset. We detected the adversarial dataset using this threshold.