Deep Steganographic Communication with Polar Code: Enhancing Robustness in Noisy Channel

초록

The reliability of secret information extraction in deep image steganography degrades severely when stego images are transmitted over open communication channels corrupted by physical noise, such as additive white Gaussian noise (AWGN). To address this challenge, this paper proposes an end-to-end deep steganographic communication framework that integrates polar codes to significantly enhance robustness under noisy transmission conditions. A key difficulty in combining channel coding with deep steganography lies in the non-differentiability of conventional encoding and decoding operations, which prevents joint optimization. To overcome this limitation, we introduce a Synergistic Training Module (STM) that enables effective co-optimization between the steganographic network and the polar coding process by exposing the deep networks to channel impairments and decoding effects during training, without requiring explicit gradient propagation through discrete coding operations. In addition, an Offset Min-Sum (OMS) based belief propagation decoder is employed to achieve a favorable trade-off between decoding accuracy and computational efficiency. Extensive experiments demonstrate that the proposed framework substantially improves transmission reliability. Under an AWGN channel, the proposed method reduces the bit error rate (BER) by multiple orders of magnitude compared with uncoded and conventionally coded baselines, achieving a BER close to 10-3 at moderate SNR levels. Meanwhile, high imperceptibility is preserved, with the stego images achieving PSNR/SSIM of 41.44 dB/0.986, and the recovered secret information maintaining high fidelity with PSNR/SSIM of 39.40 dB/0.991. These results confirm that the proposed framework achieves an effective balance between concealment and channel robustness, highlighting its strong potential for reliable covert communication over noisy open networks. © 1999-2012 IEEE.

키워드