We investigate image steganography under reencoding of images. After a secret message or a digital watermark is embedded into a jpeg image, quite frequently the image is re-encoded with a different quality factor. We model this re-encoding as a noisy communication channel with respect to the embedded message, and derive the corresponding error model. We present an error-correcting code that allows complete reconstruction of the embedded message in most cases, and partial reconstruction in all experiments performed, over a wide range of differences in quality factors.
