Self-Describing Reversible Data Hiding in Fixed-Length AMBTC Streams through Invariant Partitioning and Bounded Histogram Shifting
Year: 2026
Type: preprint
Abstract: A compressed image may have room for a message without having room for the instructions needed to recover it. This distinction is especially consequential in reversible data hiding: a cover-dependent embedding parameter can improve apparent capacity while making the original compressed stream impossible to reconstruct. We study this problem for a fixed-length, byte-addressable representation of absolute moment block truncation coding (AMBTC). The proposed scheme uses invariant low-quantizer values to identify metadata carriers and, when needed, a residue-defined subset of image blocks. A 19-bit in-band descriptor selects global or grouped embedding, an expansion peak, and a vacant quantizer bin. The overwritten descriptor-carrier bits are carried in the embedded stream and recovered before the original compressed bytes are reconstructed. We prove exact reversibility on an explicitly characterized admissible domain, unchanged stream length, and a one-gray-level bound on reconstructed-pixel perturbations. On five grayscale test images, the hybrid accepts all five covers and carries 2,049 net bits (0.00260 bits per pixel in aggregate), compared with 1,527 bits for the best single public peak shared across the corpus, selected from 254 possibilities. The aggregate difference is 34.2%, although the common-peak method accepts only four covers. At a decoded-image squared-error budget of 1,000, the hybrid carries between 12 and 48 bits per image. The scheme has a structural rate ceiling below one bit per 16-pixel block and is not shown to outperform contemporary variable-length AMBTC data-hiding methods. An encoder-reachable adversarial family additionally separates grouped and global one-peak capacities, whereas a reversible two-peak global comparator reverses that ranking on a second family. In an exploratory 25-cover extension, 75 randomized recoveries succeed, but grouping adds admissibility rather than capacity on the 21 mutually admitted covers. These results identify both the benefit and the cost of making compressed-domain reversibility self-contained.
Author: Akhtar, Mohammad Amir Khusru
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225374292