Pemanfaatan Deep Convolutional Auto-encoder untuk Mitigasi Serangan Adversarial Attack pada Citra Digital

Putu Widiarsa Kurniawan S (Institut Sains dan Teknologi Terpadu Surabaya)
Yosi Kristian (Institut Sains dan Teknologi Terpadu Surabaya)
Joan Santoso (Institut Sains dan Teknologi Terpadu Surabaya)

Journal of Information Technology
Vol 11 No 1 (2023): J-Intech : Journal of Information and Technology

Adversarial attacks on digital images pose a serious threat to the utilization of machine learning technology in various real-life applications. The Fast Gradient Sign Method (FGSM) technique has proven to be effective in conducting attacks on machine learning models, including digital images found in the ImageNet dataset. This research aims to address this issue by utilizing the Deep Convolutional Auto-encoder (AE) technique as a method for mitigating adversarial attacks on digital images.The results of the study demonstrate that FGSM attacks can be performed on the majority of digital images, although there are certain images that are more resilient to such attacks. Furthermore, the AE mitigation technique proves to be effective in reducing the impact of adversarial attacks on most digital images. The accuracy of the attack and mitigation models is measured at 14.58% and 91.67%, respectively.