Joint Image Data Hiding and Rate-Distortion Optimization in Neural Compressed Latent Representations

We present an end-to-end learned image data hiding framework that embeds and extracts secrets in the latent representations of a neural compressor. The message encoder/decoder design is flexible to support multiples of 64-bit message lengths. We jointly optimize with the rate-distortion function of the neural codec, which reduces the embedded size overhead significantly. By leveraging a perceptual loss function, our approach simultaneously achieves high image quality and low bit-error rate. Our approach offers superior image secrecy in steganography and watermarking scenarios than existing techniques. Processing messages in the compressed domain has much lower complexity, and our method can achieve about 30 times acceleration. Furthermore, with the prevalence of IoT smart devices, machines can extract hidden data directly from the compressed domain without decoding. Our framework can benefit both secure communication and the coding-for-machines concept.