ProxyPrints: Reversible Data Embedding

• ProxyPrints is a framework that embeds auxiliary data into standard images through fully reversible transformations to maintain workflow compatibility.
• It utilizes deterministic methods such as histogram shifting and secret-key rotations to ensure precise extraction of embedded layers without degrading visual quality.
• Practical evaluations demonstrate high fidelity in image printing and robust biometric matching with effective breach detection and key revocation.

ProxyPrints denotes frameworks designed for data representation that is directly compatible with existing processing or rendering workflows, yet embeds additional information such that, through a reversible transformation and extraction process, all original and auxiliary metadata can be restored. The term arises in two distinct but technologically analogous research areas: (1) secured, cancellable biometric aliasing in fingerprint systems, and (2) reversible data hiding for color and special-ink printing. Though differing in application domains, both ProxyPrints frameworks implement a separation of data layers and employ invertible transformations to guarantee both compatibility and recoverability (Hiraoka et al., 2021, Hacmon et al., 16 Nov 2025).

1. Definition and Core Principles

A ProxyPrints system deploys a deterministic, fully reversible transformation that encodes auxiliary ("special" or "alias") data within a host format (such as standard image or fingerprint scan data), retaining operational compatibility for default use-cases while permitting precise recovery of auxiliary layers when supplied with appropriate keys or extraction algorithms. In color printing, this enables a single composite image indistinguishable from conventional output but carrying embedded layers for special inks (Hiraoka et al., 2021). In fingerprint biometrics, ProxyPrints allows transforming raw fingerprint images into synthetic aliases that preserve matcher performance but are unlinkable, revocable, and support breach detection (Hacmon et al., 16 Nov 2025).

2. Methodologies and System Architectures

2.1 Special-Ink Printing via Reversible Data Hiding

Input consists of a general color image $G$ (24-bit RGB, intended for standard CMYK) and one or more special layers $S$ (e.g., white-ink mask, metallic-ink densities). Using histogram-shifting (HS) reversible data hiding (RDH), compressed representations of $S$ (by JBIG2) are embedded into select color channels—red for binary, blue for 3-bit layers—without altering overall visual appearance. Extraction precisely recovers $G$ and $S$ by inverse HS, decompression, and auxiliary side channel data recovery (Hiraoka et al., 2021).

2.2 Biometric ProxyPrints via Encoder–Rotation–Decoder

The biometric ProxyPrints pipeline receives input $x_p$ (raw fingerprint), applies a learned encoder $\mathrm{En}$ to $x_p$ mapping it onto the unit hypersphere, then performs a secret-key-dependent rotation $R_k$ in embedding space, and finally decodes via $\mathrm{De}$ to synthesize a visually realistic but deterministic alias $x_p' = T_k(x_p)$. Downstream algorithms (proprietary matcher) ingest only the alias $x_p'$ (or its minutiae-level template), ensuring original biometrics are neither exposed nor reconstructible without the secret key. Revocation, breach detection, and compatibility with matcher software are inherent architectural features (Hacmon et al., 16 Nov 2025).

Transformation Formalism in ProxyPrints

$$T_k : \mathcal{X} \longrightarrow \mathcal{X}',\quad T_k(x_p) = \mathrm{De}\bigl(R_k(\mathrm{En}(x_p))\bigr)=x_p'$$

where $R_k\in \mathrm{SO}(n)$ is a rotation parameterized by $k$.

3. Embedding and Extraction Procedures

Printing ProxyPrints–Embedding Workflow

• Preprocessing: Special layers are losslessly compressed (via JBIG2); multi-bit layers are bit-plane-split and concatenated for maximal compression.
• Histogram Shifting: For each embedding channel, identify peak $PP$ and zero (or minimal) $ZP$ bins, shift histogram bins accordingly, and embed payload bits via histogram modulation.
• Side Information: Embedding parameters $(PP, ZP, LP$ if needed$)$ are packed into image LSBs; their positions recorded to ensure reversibility.
• Capacity: Payload capacity per channel equates to the number of peak-bin pixels; effective embedding rates attained are $0.08$–$0.1$ bits per pixel.

Extraction reverses each step, ensuring lossless recovery of both $G$ and $S$ (Hiraoka et al., 2021).

Biometric ProxyPrints–Alias Generation and Matching

Enrollment and authentication comprise:

Enrollment(x_p, ID): e ← En(x_p) e' ← R_k(e) x_p' ← De(e') t' ← ExtractMinutiae(x_p') Store(DB, ID, t')

Authentication(x_p_try): e ← En(x_p_try) e' ← R_k(e) x_p'_try ← De(e') t'_try ← ExtractMinutiae(x_p'_try) For each stored (ID_i, t'_i): s_i = match(t'_i, t'_try) If max_i s_i ≥ τ: accept(ID_argmax); else: reject

Spoofed or replayed aliases fail to match after transformation and can be detected (Hacmon et al., 16 Nov 2025).

4. Performance Metrics and Experimental Results

Printing

On 830 × 1,170 illustrations, after compression:

• Payloads: $3$–$12$ KB for binary, $3$–$24$ KB for 3-bit layers.
• Distortion: Luminance PSNR ≈ ∞ dB, MSSIM ≈ $1.000$ for unaltered channels; red/blue channels PSNR $56$–$67$ dB, MSSIM $\geq 0.9965$.
• Subjective indistinguishability: Pixel-level shifts undetectable in observer 2AFC tests.
• Capacity: Sufficient for high-fidelity mask embedding in illustrations; natural images give lower capacity due to dense histograms (Hiraoka et al., 2021).

Biometrics

On LivDet 09–15 (3,516 identities, $\sim$24.7K images):

Metric	Baseline (Bozorth3)	ProxyPrints + Bozorth3
ROC AUC	0.93	0.86
PRAUC	0.95	0.87
EER	0.14	0.19
F1 @ $\tau=40$	0.79	0.70
Recall @ $\tau=40$	0.66	0.58

Detection rate for spoof/replay 98.97%; direct alias replay flagged with 99.96% accuracy. Processing overhead approximately $200$ ms/scan (Hacmon et al., 16 Nov 2025).

5. Security Properties and Theoretical Guarantees

Printing

• Perfect reversibility: All embedding/extraction steps are information lossless, provided histogram/side info fits within capacity.
• No visible artifact: Chrominance or luminance distortion statistically imperceptible under typical viewing.

Biometrics

• Determinism: Same finger, same key, same alias.
• Non-invertibility: Infeasible to reconstruct input without secret $k$.
• Revocability / Key-rotation: All aliases immediately unusable when switching $k$; practical “cancellable biometrics.”
• Unlinkability: Matches across rotated aliases always below threshold $\tau$.
• Breach detection: Attempted replay or spoof using stored aliases is unambiguously flagged (Hacmon et al., 16 Nov 2025).

6. Applications, Limitations, and Future Directions

Applications

• Printing: Single “proxy print” artifacts can flow through standard CMYK pipelines, carrying embedded special-ink layers for later high-end reproduction, reducing archival and workflow complexity. A plausible implication is that such proxy prints could be integrated into existing digital asset management systems to manage both standard and specialty print workflows without format multiplication.
• Biometrics: “Drop-in” middleware protection layer for legacy and proprietary fingerprint matchers, enabling key-rotatable, privacy-preserving biometric templates and real-time breach/spoof detection.

Limitations

• Printing: Scheme tailored for illustrations with sparse color histograms; natural photographs may require alternative (e.g., prediction-error expansion) RDH variants. Demonstrated only up to 3-bit-depth masks; higher complexity demands advanced compression/embedding.
• Biometrics: Revocation requires user re-enrollment after key change. Non-invertibility lacks a formal cryptographic proof. Robustness depends on large, heterogeneous training datasets. Advanced liveness-detection bypass attacks are out-of-scope but can be layered orthogonally (Hiraoka et al., 2021, Hacmon et al., 16 Nov 2025).

Future extensions may encompass frequency-domain RDH, multi-modal biometric aliasing, and formalization of non-invertibility properties for cryptographic assurances.

7. Comparison to Baselines

In printing, the classic histogram-shifting RDH outperforms HDWT and difference-expansion under sparse-color conditions. No embedded-data-hiding scheme prior explicitly targeted special-ink compatibility with full reversibility and no extra print channels. In biometrics, ProxyPrints is the first approach to provide fully cancellable, transparent, key-rotatable aliases without requiring matcher modifications; previous template-protection and cancellable biometric frameworks often break matcher compatibility or are susceptible to record multiplicity and inversion attacks (Hiraoka et al., 2021, Hacmon et al., 16 Nov 2025).