Defense Against Image Captioning Attacks via A Robust and Stable Recurrent Neural Network
Jiahuan Zhang, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama · 2021 IEEE 10th Global Conference on Consumer Electronics (GCCE) · 2021
This paper proposes a defense method against image captioning attacks via a robust and stable recurrent neural network (RNN). The improved RNN is constructed based on the dynamical system theory. The recurrent unit in this RNN has a very high ability to express hidden states. From the perspective of the theory, the improved RNN has global exponential stability. Then we conduct noun-type attack experiments on image captioning systems for comparing a modified vanilla RNN with the improved RNN, respectively. The experimental results fully illustrate the effectiveness and robustness of the improved RNN for the image captioning system.