Secret image sharing splits a picture into several same-size "shares" so that any $r$ of $n$ holders can rebuild it exactly, while fewer than $r$ learn nothing. The most secure recent schemes chain the shares: each pixel's sharing step borrows randomness from the previous step's output, so a one-pixel change in the secret scrambles the whole share. That chaining is a sequential pipeline, and sequential pipelines turn one small error into a large one. Flip one bit in one stored or transmitted share and the chain can corrupt everything decoded after it. The scheme is lossless only on a perfect channel.
The fix proposed here needs no extra bytes. A threshold scheme with $n > r$ already carries redundancy. Shamir-style shares at one pixel position are the symbols of a Reed–Solomon codeword, so the $n-r$ surplus shares can locate and repair corrupted ones. Chaining wastes this, because the damage from one bad pixel spreads across every later position and overwhelms the code. We propose to rebuild the chain so that each step draws its randomness from share values the receiver has already repaired, and never from values it has yet to check. Errors are then fixed position by position before they can travel downstream. Recovery stays exact (infinite PSNR), with no pixel expansion and the chain's diffusion kept.
Core Claim
If the feedback state for position $t$ is a keyed function of the error-corrected shares at earlier positions, then decoding runs as a loop. At each position, the receiver first repairs the shares with a cross-share Reed–Solomon decoder, then updates the feedback state from the repaired values. Error propagation then drops to zero whenever each position has at most $\lfloor (n-r)/2 \rfloor$ wrong shares, or at most $n-r$ shares known to be missing. In the language of On Error Propagation of Diffusion Models, the amplification factor becomes 0 inside the correction radius. Outside it, the damage should stay within a bounded window rather than growing without limit.
What would be new is a chained, dimension-preserving image sharing scheme with two properties at once. It keeps near-optimal differential-attack resistance, measured by NPCR and UACI: the fraction of share pixels that change, and how much they change, when one secret pixel is altered. And it gives a provable bound on how far channel errors spread during decoding. The cost is plain. Security at the chained positions rests on the secrecy of a key, which is computational rather than information-theoretic, and correction capacity uses up surplus shares that could otherwise just sit idle.
Where This Sits
This proposal builds most directly on Reversible Data Hiding in Encrypted Images Using Cipher-Feedback Secret Sharing. That paper chains Shamir-style sharing in the manner of block-cipher feedback modes to resist differential and chosen-plaintext attacks. It reports lossless reconstruction and embedding rates of 1.25–3.64 bpp on the standard 512×512 set. It does not study channel errors. Whether its feedback is fed by plaintext, by intermediate coefficients, or by share values determines whether an error keeps spreading forever or stops after one step. Our first task is to establish that experimentally.
Reversible Data Hiding over Encrypted Images via Intrinsic Correlation in Block-Based Secret Sharing supplies the second vulnerable chain and the unchained comparison. Its size-reduced variant discards pixel shares and arithmetic-codes MED prediction errors; MED is a median-edge-detector predictor that guesses each pixel from its neighbours. An arithmetic decoder loses its place after one corrupted bit, so any bit error in that stream becomes a global failure. Algebraic schemes based on the Chinese Remainder Theorem and Boolean XOR sharing motivate the zero-expansion constraint. They also show the fragility we are addressing: any corrupted residue makes exact inversion fail.
The sequential-error literature is our lens and our constraint rather than our method. Restoration work reaches the same diagnosis again and again: downstream stages treat upstream errors as real signal. Deep Residual Network for Joint Demosaicing and Super-Resolution and Gated Fusion Network for Joint Image Deblurring and Super-Resolution respond by merging the stages into one network. That answer is unavailable to us, because merged networks are approximate. On Error Propagation of Diffusion Models formalizes propagation through a per-step amplification factor, which it shows is at least 1 for standard diffusion. It also shows that input perturbation does not fix this. We borrow that measurement, since a discrete chain allows the factor to be measured exactly. Our mechanism differs from both responses: we keep the stages separate and put an exact corrector between them.
Generative reversible-privacy systems cap out below exact recovery. HIPP reports SSIM 0.9899, and Diff-Privacy reports 28.9–31.9 dB. These show why a learned approach cannot deliver infinite PSNR. We do not try to make shares look natural.
Research Program
The first stage is a measurement study. We will reimplement the chained sharing of Hua et al. and the size-reduced block scheme of Zou et al. We will then inject controlled faults at the share level: single bit flips, random flips at a range of rates, burst errors, dropped rows (packet erasures), and one entire share corrupted. For each fault we record the spread length, meaning how many decoded pixels differ from the secret, and the empirical amplification factor, meaning the ratio of corrupted positions at step $t+1$ to those at step $t$. This alone fills a gap, since neither paper reports robustness.
The second stage is the construction. We work over GF($2^8$) so that 8-bit pixels fit exactly, with evaluation points $x_1,\dots,x_n$. At position $t$ the share for holder $i$ is
$$s_i(t) = m(t) + \sum_{j=1}^{r-1} a_j(t)\, x_i^{\,j}.$$
Here $m(t)$ is the secret pixel. The random coefficients are $a_j(t) = F_K\big(\hat{s}(t-w{:}t-1)\big)_j$, where $F_K$ is a keyed pseudorandom function and $\hat{s}$ are the corrected share vectors from the previous $w$ positions.
The encoder computes the same $\hat{s}$ it will transmit, so the chain is deterministic. The decoder runs the same loop in order. At each position it applies Berlekamp–Welch decoding (a standard Reed–Solomon decoder that finds and fixes wrong symbols) to the $n$ received symbols, reads off $m(t)$, then feeds the repaired vector forward. The pixel-to-pixel diffusion comes from feeding the chain with shares that already depend on $m(t-1)$, together with an optional reverse second pass. We will prove two results. The first is zero propagation inside the correction radius. The second is a spread bound of $w$ positions per uncorrectable event, which holds whenever the decoder can detect failure and flag the affected positions as erasures.
The third stage makes the data-hiding layer survive the same faults. The embedded payload and the location maps are the brittle part: in Zou et al., side information is arithmetic-coded. We will split that stream into independently decodable segments aligned to image blocks. The loss in bpp from segmentation will be measured, not assumed. The reserved-bit cost should come in below adding parity to each share.
Baselines
The naive reference is the unmodified chained scheme of Hua et al. under the same fault models. It reports lossless recovery on clean channels and embedding of 1.25–3.64 bpp. No reported result found on Lacuna for its behaviour under channel errors. Any exact-recovery rate above the near-zero we expect for this baseline is informative, but the baseline itself must be measured.
The strongest competitor is unchained block Shamir sharing, as in Zou et al., combined with cross-share Reed–Solomon decoding. Without a chain there is nothing to propagate, so this baseline should match our robustness exactly. Its expected weakness is differential resistance. Zou et al. report 2.7798 bpp on BOSSBase and 2.6222 bpp on BOWS-2 for their size-reduced scheme. They report no NPCR or UACI figures on Lacuna, and Hua et al. report NPCR and UACI only as "close to theoretical optima". So both schemes' diffusion scores will be recomputed on identical images.
We beat this competitor only if two things hold. Our exact-recovery rate under faults must be statistically indistinguishable from it. And our NPCR and UACI must sit near the ideal values (about 99.6% and 33.5% for 8-bit data), while its scores fall clearly short.
A third baseline appends a systematic Reed–Solomon parity block to each chained share at matched redundancy. This makes the expansion cost explicit. At matched expansion it is a fair fight. At zero expansion it cannot exist.
Two ablations isolate the key ingredient. In the first, the chain is fed from uncorrected received values, with cross-share decoding still applied. This tests whether "correct, then feed forward" is what matters. In the second, the chain is corrected but its window $w$ is enlarged, to trace the robustness–diffusion trade-off.
All methods share the same images, the same $(r,n)$ values in $\{(2,4),(3,5),(3,7)\}$, the same fault seeds, and the same field implementation. Runtime is measured on one CPU core.
Why not just…?
Why not just drop the chain and use plain Shamir with Reed–Solomon decoding? This is the real threat. If the unchained scheme's diffusion scores prove adequate against the differential attacks Hua et al. target, the chain adds nothing and our contribution shrinks to a measurement study. The matched NPCR, UACI and chosen-plaintext comparison above settles this. We commit to reporting that outcome whichever way it falls.
Why not just add error-correcting parity to each share? Parity bytes inside a single share protect that share's bytes, but they enlarge the share, which breaks the zero-expansion property. Alternatively they occupy vacated room that Hua et al. and Zou et al. give to payload, at 1.25–3.64 and 2.62–6.56 bpp respectively. Our scheme reuses redundancy the threshold structure already contains. The matched-redundancy baseline measures what per-share parity buys on top of that and what it costs.
Why not train one network end to end, as restoration work does? Merged networks avoid the stage-to-stage cascade but deliver approximate output. The best figures in the file are 31.41 dB for demosaicing plus super-resolution, SSIM 0.9899 for HIPP, and at most 31.9 dB for Diff-Privacy. None of these reaches exact recovery, and none carries a threshold-security guarantee. We do not run this baseline.
Why not smooth or filter between stages? Shares are pseudorandom field elements, so any filtering corrupts them outright. Li and van der Schaar also show that perturbation-style fixes leave amplification in place. For the same reason we scope out lossy compression of shares. We will report only where mild JPEG compression breaks each scheme, and we do not claim robustness to it.
Proposed Experiments
Experiments run on the eight standard 512×512 images for per-image security metrics, and on BOSSBase and BOWS-2 (10,000 images each) for capacity and robustness statistics.
The primary metric is the exact-recovery rate: the fraction of images decoded with infinite PSNR and SSIM of 1.0. Secondary metrics are PSNR when recovery is not exact, spread length, the empirical amplification factor, NPCR, UACI, the correlation between neighbouring share pixels, embedding rate in bpp, and expansion ratio.
The decisive check is a sweep of symbol error rate for $(3,7)$. Theory predicts a sharp threshold where more than two wrong shares at one position becomes likely. Below that threshold our scheme and the unchained Reed–Solomon baseline should both stay at exact recovery, while the naive chained scheme collapses after its first error. Above it, our spread length should stay near $w$ rather than growing with image size.
The second decisive check is the first ablation, chain fed from uncorrected values. If it matches the full method, the "correct, then feed forward" ordering is not the reason for the gain and the core claim fails. The third check is the NPCR and UACI comparison against unchained sharing, which decides whether the chain earns its place. Multi-hider runs will embed payload from several data hiders, including one that writes faulty bits, and report the bpp lost to segmentation of the side-information stream.
References
- On Error Propagation of Diffusion Models
- Reversible Data Hiding in Encrypted Images Using Cipher-Feedback Secret Sharing
- Reversible Data Hiding over Encrypted Images via Intrinsic Correlation in Block-Based Secret Sharing
- Secret Image Sharing Based on Chaotic Map and Chinese Remainder Theorem
- An Adaptable (n, n) Secret Image Sharing Mechanism Based on Boolean Operation
- Deep Residual Network for Joint Demosaicing and Super-Resolution
- Gated Fusion Network for Joint Image Deblurring and Super-Resolution
- HIPP: Protecting Image Privacy via High-Quality Reversible Protected Version
- Diff-Privacy: Diffusion-Based Face Privacy Protection