Neural Network Based Watermarking Trained with Quantized Activation Function

We propose a watermarking method that incor-porates a quantized activation function to provide robustness against quantization. Zhu et al. showed that the introduction of a noise layer between the encoder and decoder can increase the robustness against attacks. Although there are various attacks on stego-images, these images are often JPEG-compressed. As the process of JPEG compression includes quantization, the watermark decoder must be able to estimate watermarks from compressed images. Hence, we propose a quantization layer that introduces a quantized activation function consisting of the hy-perbolic tangent function. The proposed neural network is based on that proposed by Hamamoto and Kawamura. By simulating the quantization of JPEG compression, the quantization layer is expected to improve the robustness against JPEG compression. The robustness was evaluated by the bit error rate (BER), and the stego-image quality was evaluated by the peak signal-to-noise ratio (PSNR). The proposed network achieved a high image quality of more than 35 dB, and it could extract watermarks with a BER of less than 0.1 for Q-values of 30 or higher in JPEG compression. It was thus more robust against JPEG compression than Hamamoto and Kawamura's model.