Latent Space Encryption with PWLCM for Enhanced Secure Deep Joint Source Channel Coding
Yuyang Fu, Katsuya Suto · 2024
In recent years, research and development of deep learning based joint source channel coding (DJSCC) has become increasingly prevalent. Compared to traditional separate source channel coding (SSCC) schemes, DJSCC demonstrates exceptional performance, particularly in environments with low signal-to-noise ratios (SNR) and limited bandwidth. However, the protection of private information during transmission remains a critical concern. A notable weakness of DJSCC is its incompatibility with traditional encryption methods used for secure communications, making it vulnerable to eavesdropping attacks. We propose the integration of a chaotic image encryption method into the DJSCC model for secure wireless image transmission. We encrypt the latent space, allowing encryption and DJSCC to be designed individually. Our proposed method not only ensures the receiver maintains the same transmission characteristics as traditional DJSCC and achieves high reconstructed image quality but also effectively prevents eavesdropping. Furthermore, the method’s high key-sensitive enables it to have robustness against differential attacks.