.png){kind=icon}
.png){kind=icon}
Return
Robust steganography for convolutional neural networks
.png){kind=icon}
PRE
AI
DOI：10.1117/1.JEI.34.6.063002.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, propelled by the remarkable progress in deep learning technologies, model steganography based on deep neural networks has become a significant focus of academic research. Unlike traditional steganographic methods, deep-model steganography excels in its superior imperceptibility. Current deep model steganography methods generally embed secret data in the intricate weight parameters of neural networks for covert transmission. However, these approaches tend to significantly affect the host neural network's primary task performance and are susceptible to high-rate pruning attacks. We introduce a loss function designed to mitigate the negative impacts on the original model's performance when embedding secret information in convolutional layer weights. In addition, we propose an improved adaptive embedding method to counter common pruning attacks. Experimental results show that the proposed method effectively enhances both the embedding capacity of network weights and the security of steganography. Remarkably, even with a drastic reduction in model parameters, our approach remains robust against pruning attacks.
Keywords:
steganography
neural networks
covert transmission
robustness
Journal
J
IF:
1
Papers:
148
Citations:
2.7K