http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
부분적인 스크린 영상 혼합을 통한 합성곱 신경망의 영상 인식 성능 향상
변준영(Junyoung Byun),심규진(Kyujin Shim),김창익(Changick Kim) 대한전자공학회 2019 대한전자공학회 학술대회 Vol.2019 No.11
Data augmentation is a way of improving the generalization ability of deep neural networks by expanding their training data with pre-defined transformations. For image recognition tasks, traditional data augmentation methods transform an image by using simple techniques such as random horizontal flipping and cropping. However, several recent methods have been proposed to augment training data by linearly interpolating two images with arbitrary proportions. Although they can further improve the generalization ability, they use a randomly chosen mixing ratio throughout a pair of images, which may ignore their local characteristics. In this paper, we propose a novel data augmentation method that can vary the mixing ratio according to the local brightness. Our method partially blends two images with Screen blend mode. We have also shown that CNNs can be successfully trained, only with the blended inputs. Experimental results on the CIFAR-10 and CIFAR-100 datasets have shown that the proposed method yields superior performance than existing methods.
전이 기반 적대적 공격 방어를 위한 신경망의 특성 지도 데이터 무작위 치환
김정수(Jeongsoo Kim),변준영(Junyoung Byun),김창익(Changick Kim) 대한전자공학회 2020 대한전자공학회 학술대회 Vol.2020 No.11
Although convolutional neural networks show good performance in the field of image recognition, adversarial attacks on them have become a big threat in recent years because it can cause the neural networks to malfunction by intentionally adding small noise. Such adversarial examples generated from the attacker’s model can also deceive other networks and this intriguing property is called transferability. The more similar the architecture of the two models and the training data, the higher the probability of the attack of the attack being valid. However, it is difficult to change the structure of each neural network dramatically to prevent transferability-based attacks. In this paper, we propose a method to make a model to have a unique computational process by mixing the input feature maps of the convolutional layer without changing the structure of the neural network in order to improve the robustness against transferability-based adversarial attacks. We demonstrate the effectiveness of our method with various convolutional neural networks trained on the CIFAR-10 dataset.