WorksComputer Science
Attack-Resilient Multi-Bit Image Watermarking Framework via Latent Coordinate Sign Clamping
Abstract
Watermarking is essential for image copyright protection and attribution, yet existing methods suffer from fundamental trade-offs among post-hoc applicability to pre-existing images, multi-bit capacity, training-free deployment, and adversarial resilience. To address these limitations, we propose a training-free watermarking framework that exploits the overcompleteness of diffusion latent spaces to embed multi-bit payloads into arbitrary images while preserving high visual quality. Additionally, we develop a three-layer decoding pipeline that combines magnitude-weighted majority voting, BCH, and Chase decoding to maximize message-extraction accuracy in the presence of channel noise. To counter active adversaries, our framework integrates a steganalysis-resistant noise camouflage layer and a dual-layer forgery defense that combines perceptual hashing with a Reed-Solomon fuzzy commitment scheme over CLIP features. Empirical evaluations demonstrate that our method applies to both natural and AI-generated images with a scalable -bit capacity. Furthermore, it effectively mitigates advanced removal vectors ( decode rate against CtrlRegen+ regeneration attacks) and secures robust forgery detection ( accuracy via joint CLIP and RS fuzzy commitment mechanism).
Introduction
While generative AI democratizes high-fidelity content creation, it simultaneously accelerates the spread of misinformation via deepfakes and undermines digital asset attribution [4, 16] (Yu et al. 2024). To restore ecosystem trust and enforce intellectual property protection, digital watermarking has emerged as a vital paradigm for end-to-end provenance verification [34, 3]. As text-to-image models scale, embedding robust, multi-bit watermarks as intrinsic audit trails provides a critical forensic mechanism, enabling downstream defenders to resolve the exact model, user, or platform responsible for an asset’s lineage.
Despite significant progress, existing image watermarking methods face a strict four-way trade-off among post-hoc applicability to existing images, multi-bit capacity, training-free deployment, and adversarial resilience. Post-processing methods [1, 15, 35, 23, 10, 31] directly modify pixel space. While fast and robust to benign transformations, they are easily neutralized by generative regeneration attacks [33]. Conversely, in-processing techniques [25, 8, 28, 27] (Yang et al. 2024) integrate watermarks directly into the content generation pipeline, but they cannot accommodate pre-existing images and require intrusive model modifications or fine-tuning. Specifically, ZoDiac [32] leverages latent optimization to bridge the existing-image gap, but it is restricted to zero-bit detection, preventing granular user attribution. Similarly, MetaSeal (Zhou et al. 2026) uses cryptographic watermarks to thwart forgery, yet its reliance on Invertible Neural Networks leaves it critically vulnerable to regeneration attacks. Crucially, no state-of-the-art paradigm simultaneously secures image attribution against both watermark removal and malicious forgery.
To address these challenges, this paper introduces a training-free, robust, multi-bit watermarking framework tailored for pre-existing imagery. Figure 1 demonstrates the high-level architecture of our proposed framework. The watermark embedding pipeline maps the source image to its noisy latent via DDIM (Denoising Diffusion Implicit Models) inversion [20] and embeds the payload by sign-encoding high-magnitude values into a keyed subset of coordinates, optimized via adaptive -mixing to preserve quality. We assume an adversary can intercept these watermarked images from the public network to execute watermark removal or forgery attacks. To counter this, the verification pipeline inverts the suspect image to and extracts raw bits from the designated coordinates. These are processed through a joint decoder (majority voting, BCH, and Chase decoding) to reconstruct the payload, which is validated against cryptographic content-binding metadata.

Figure 1. The High-level architecture. Left-top: The platform embeds a user-specific watermark (ID with secret key ) into the original image before distribution. Left-bottom: The image is leaked or redistributed by an unauthorized adversary through diverse attacks. Right: Given suspicious images, the platform detects the watermark to identify the user and verifies image content authenticity.
As illustrated by the comparative analysis in Table 1, our framework introduces a unified architecture that successfully reconciles the trade-offs inherent in competing watermarking paradigms through four core innovations. First, by utilizing DDIM inversion to remap the latent vector and iteratively optimize the initial noise map of a pre-trained, frozen LDM (Latent Diffusion Model), we achieve an entirely training-free deployment that routinely outperforms traditional pre-training and fine-tuning baselines. Second, the framework is post-hoc applicable. It embeds multi-bit payloads into the latent space of arbitrary natural or synthetic images. This expanded capacity allows the watermark to convey the rich metadata required for definitive source attribution. Third, by operating directly in the latent space rather than altering raw pixel values, the embedded payloads remain perceptually unnoticeable while maintaining resilience to adversarial conditions, most notably by resisting Stable Diffusion-based regeneration attacks. Finally, our verification paradigm integrates content-binding with a cryptographic commitment framework rather than relying on simple presence detection, thereby systematically mitigating the risk of adversarial watermark forgery and replay attacks.
Our main contributions are:
We propose a training- and fine-tuning-free framework that embeds multi-bit identifiers (32–128 bits) into arbitrary pre-existing images. The approach demonstrates robust resilience against severe distortions, achieving a 94.6% exact 32-bit decoding rate under diffusion-based regeneration attacks.
We demonstrate that clamping of diffusion latent dimensions (16,384 for a 512512 image) structurally protects the watermark without disrupting reconstruction quality SSIM . We also expose a non-monotonic interplay between encoding amplitude and -mixing, establishing an empirical operating threshold () that balances signal gain and distortion.
We develop a three-layer decoding pipeline (magnitude-weighted majority voting, BCH, and Chase decoding [5]) that boosts extraction accuracy from 81.8% to 91.7% over noisy diffusion channels. This architecture is coupled with a steganalysis-resistant noise camouflage layer (mean KS ) and a dual-layer forgery defense that combines perceptual hashing with a cryptographic fuzzy commitment scheme over CLIP [17] features, successfully thwarting up to 98.8% of watermark transplantation attacks.
Related Work
Latent-Space Watermarking Methods
Latent-space watermarking preserves image fidelity by embedding payloads directly into the intermediate representations or initial noise of generative models. Tree-Ring [25] embeds Fourier-domain patterns in the initial noise for zero-bit detection via DDIM inversion, and RingID (Ci et al. 2024) extends this to multi-key identification by assigning distinct ring patterns to different users. Gaussian Shading [29] utilizes key-dependent Gaussian noise patterns. However, these methods embed watermarks only during generation and cannot be applied to pre-existing images. ZoDiac [32] bridges this gap by applying static concentric Fourier rings to pre-existing images via latent optimization, but remains a zero-bit scheme incapable of user attribution.
To embed multi-bit payloads, in-processing methods like PRC [8], Xiong et al. [28], and OptMark [27] inject tracking data via pseudorandom sign-encoding or multi-step latent perturbations. Although resilient to regeneration attack, they cannot support pre-existing images and require model fine-tuning or architectural modifications. Similarly, latent-embedding schemes like DistSeal [18] and RAIN [22] avoid structural changes but necessitate intensive post-hoc training to guarantee encoder-decoder consistency and reliable watermark recovery.
Our framework bridges this architectural gap by porting the sign-encoding paradigm of PRC into ZoDiac’s image optimization pipeline without training an encoder-decoder pair. Crucially, we substitute ZoDiac’s static ring geometry with user-specific signed coordinates in the spatial domain of . To optimize error correction over the diffusion channel, we deploy a concatenated BCH code paired with Chase decoding [5] rather than PRC’s traditional LDPC architecture (See § 3).
Forgery Defense and Content Binding
Latent-space watermarking is highly susceptible to forgery (replay) attacks, wherein adversaries transfer static watermark signatures (e.g, Tree-Ring and Gaussian Shading) onto unwatermarked assets without knowledge of the underlying message [14, 9]. Specifically, Müller et al. (2025) utilize cross-architecture latent manipulation to imprint target watermarks, while Jain et al. [9] exploit the many-to-one pixel-to-latent mapping to craft black-box perturbations from a single reference image. Beyond fixed-pattern exploits, sophisticated adversaries can bypass naive protection by developing adversarial hash collisions between visually disparate media [21]. To counter this, content-dependent watermarking binds payloads directly to image semantics. While frameworks like MetaSeal (Zhou et al. 2026) enforce semantic cryptographic signatures, they require resource-intensive encoder-decoder training. In contrast, our framework introduces a training-free content-binding mechanism: we jointly encode a perceptual hash with the user ID and apply a cryptographic fuzzy commitment scheme [11] over CLIP features to secure the authentication pipeline against adversarial collision vectors.
Methodology
Threat Model
Assumptions Adhering to Kerckhoffs’s principle, we assume the adversary possesses comprehensive knowledge of the watermarking framework—including embedding and detection procedures—and access to watermarked samples, while the user’s private cryptographic key remains strictly confidential.
Adversary Models We evaluate two distinct adversarial objectives: (1) Removal attackers aim to degrade or erase the embedded watermark to prevent user-ID recovery. They may deploy standard transformations (JPEG, noise, blur, contrast adjustments), neural image compression [6],
or diffusion-based regeneration attacks [13]. (2) Forgery attackers attempt to transplant a watermark from a source asset into an unrelated target image, exploiting the fixed-pattern vulnerability of zero-bit schemes to trigger false provenance attribution.
Protector Objectives The protector (content owner or hosting platform) maintains white-box access to the diffusion model and holds the secret cryptographic keys for all registered users. The protector embeds a user-specific signature prior to image distribution. Upon auditing a suspect image, the detector extracts the payload to fulfill a dual objective: (1) attribution (reliably identifying the distributing user) and (2) forgery detection (determining whether the watermark was maliciously transplanted).
Overview
Given a source image and a user identifier , our framework generates a watermarked asset that preserves structural fidelity (SSIM ) while guaranteeing accurate payload recovery under adversarial degradation. As illustrated in Figure 2, the embedding pipeline proceeds through three core stages: (1) DDIM inversion to extract the initial noisy latent , (2) iterative optimization of with amplified sign encoding re-injected at cryptographically keyed coordinates every iteration, and (3) -mixing with the source image to meet the target SSIM. The verification pipeline mirrors this trajectory: it applies DDIM inversion to a suspect image, extracts the latent sign bits at the designated coordinates, and reconstructs the user ID via a joint decoding pipeline comprising majority voting, BCH error correction, and Chase decoding. To fortify the security boundary, the core architecture is augmented with per-user key-based coordinate isolation, noise camouflage to resist statistical steganalysis, and a dual-layer content-binding mechanism (perceptual hashing and cryptographic fuzzy commitment) to thwart forgery attacks. The primary mathematical notations are summarized in Table 1 from the supplementary material.
| Method | Existing images | Multi-bit watermark | Training-free | Attack resilience |
| HiDDeN | ✓ | ✓ | × | × |
| StegaStamp | ✓ | ✓ | × | × |
| Tree-Ring | × | × | ✓ | ◐ |
| PRC | × | ✓ | ✓ | ◐ |
| Xiong et al. | × | ✓ | × | ◐ |
| OptMark | × | ✓ | ✓ | ◐ |
| ZoDiac | ✓ | × | ✓ | ◐ |
| MetaSeal | ✓ | ✓ | × | ◐ |
| Ours | ✓ | ✓ | ✓ | ✓ |
Table 1. ✓ = Supported ◐ = Partial × = Not supported
Figure 2. Overview of the embedding (top) and detection (bottom) pipelines. During embedding, the input image is DDIM-inverted, watermark signs are injected at secret with amplitude , and is iteratively optimized to preserve image quality. The watermarked image is -mixed with the original to achieve the target SSIM. During detection, DDIM inversion is performed once; each candidate user ID requires only unscrambling, majority voting, and BCH decoding. The diffusion model remains frozen throughout.
Embedding Pipeline
Algorithm 1 describes the embedding procedure. The latent vector corresponding to image is obtained via DDIM inversion, denoted as , where is a pre-trained VAE (Variational Autoencoder) encoder. The pipeline encodes a user ID into the signs of a small subset of latent coordinates. Each user holds an independent 256-bit key , from which watermark positions are derived via SHA-256. The 32-bit user ID is BCH(63,36,)-encoded into 63 coded bits, each repeated 5 times and mapped to , yielding a codeword after per-user pseudorandom scrambling. Each selected coordinate is then set to (amplitude ), modifying only of the -dimensional latent space. The VAE decoder then reconstructs the watermarked image from the DDIM-denoised latent, prior to optimization.
Throughout this optimization process, the weights of the diffusion model remain frozen, while gradients are backpropagated exclusively to update the latent representation . Similar to ZoDiac (Zhang et al. [32]), the reconstruction objective is formulated as a joint loss function that combines pixel-level Mean Squared Error (MSE), the Structural Similarity Index Measure (SSIM), and Watson’s perceptual loss. Consequently, the computational overhead of the embedding phase is fundamentally dominated by the diffusion forward passes required during the optimization loop. We provide comprehensive details on DDIM inversion and the encoding process in Sections 1 and 4 of the supplementary material.
| Algorithm 1: Watermark Embedding |
| Require: Image , key , user ID , amplitude , SSIM threshold , iterations |
| Ensure: Watermarked image |
| 1: |
| 2: |
| 3: repeat times |
| 4: for all |
| 5: |
| 6: Update via |
| 7: until iterations |
| 8: for all ▷ final re-injection |
| 9: |
| 10: s.t. |
| 11: return |
Table 1 (PDF p. 4).
Detection Pipeline
Algorithm 2 describes the detection procedure. The most computationally intensive operation—the DDIM inversion—is executed exactly once, independent of the number of candidate user identifiers being evaluated. For each candidate user validation, the procedure merely extracts values from the already-recovered latent vector , performs a magnitude-weighted majority vote over each group of repetitions to reconstruct the coded bits, and invokes BCH and Chase decoding. Compared to the initial inversion, these subsequent evaluation steps incur negligible computational overhead. As a result, for a database containing registered users, the overall detection complexity scales as , where the inversion latency asymptotically dominates.
Chase-BCH Decoding
Standard BCH decoding fails when more than coded bits are wrong. However, the magnitude-weighted majority vote produces not only a hard decision per coded bit but also a confidence value—the absolute value of the weighted sum. Low confidence indicates the group of repeats disagreed, suggesting the bit is likely wrong. Chase decoding (Chase [5]) exploits this: it identifies the least-confident coded bits, tries all combinations of flipping them, and attempts BCH decoding on each. If any combination reduces the error count to , decoding succeeds. With , the decoder tries 128 combinations. In the best case, this corrects up to errors, though the actual gain depends on the correlation between error positions and low-confidence bits.
| Algorithm 2: Watermark Detection |
| Require: Test image , key , candidate user ID , message bits |
| Ensure: Match |
| 1: ▷ read positions |
| 2: for |
| 3: Unscramble using seed from |
| 4: for to do |
| 5: ▷ weighted vote |
| 6: ; |
| 7: end for |
| 8: |
| 9: |
| 10: return |
Table 2.
Steganalysis Defense: Noise Camouflage
Amplifying coordinates to magnitude alters the statistical distribution of . An attacker who can DDIM-invert the watermarked image may detect the positions by identifying coordinates with unusually high recovered magnitudes. To counter this, we introduce noise camouflage: in addition to the real watermark positions, we amplify decoy positions to the same magnitude with random signs. The attacker observes loud coordinates but cannot determine which carry the message. The detector, which holds the key, knows exactly which positions are real and ignores the decoys during decoding. Critically, decoys must use the same amplitude as real positions; using a lower amplitude allows the attacker to separate them by magnitude. The steganalysis defense is evaluated in Section 6.4 of the supplementary material.
Forgery Defense: Content-Binding
Our multi-bit watermark capacity enables defense against forgery by encoding a content-bound message—the user ID plus an image-specific fingerprint—within the same 317 positions. The system supports two modes: tracking mode (32-bit ID only, maximum robustness) and anti-forgery mode (32-bit ID + 16-bit content fingerprint via unequal error protection). In anti-forgery mode, the user ID uses BCH(63,36,) with 3 repeats per coded bit (189 positions), while the 16-bit fingerprint uses 8 repeats per bit (128 positions). At detection, the decoder recovers both components and recomputes the fingerprint from the test image; a mismatch reveals forgery.
We provide two content fingerprinting options: (1) a perceptual hash (pHash) for practical detection (441/500 forgeries caught), and (2) a cryptographic fuzzy commitment (Juels and Wattenberg [11]) over CLIP features (Radford et al. [17]) that resists adversarial hash manipulation (98.8% detection). The fuzzy commitment derives a content-binding key from a Reed–Solomon-protected secret locked to the image’s CLIP features; forgery causes decoding failure and key mismatch. Full details are in § 5 of the supplementary material.
Experiments
Experimental Setup
Datasets and models We evaluate our framework across three distinct benchmarks, with all assets standardized to a resolution of : (1) MS-COCO val2017 (Lin et al. [12]), comprising 500 real-world photographs unexposed to generative models; (2) DiffusionDB (Wang et al. [24]), consisting of 500 synthetic images generated from open-ended text prompts; and (3) WikiArt (WikiArt [26]), containing 500 art-style images synthesized from style-specific prompts. Synthetic generations for both DiffusionDB and WikiArt are executed via Stable Diffusion 2.1.
Implementation Details All experiments are implemented in PyTorch and executed on a single NVIDIA RTX 5090 GPU. Our core setup integrates Stable Diffusion 2.1 configured with 50 DDIM denoising steps, latent coordinates, encoding amplitude (Section 6.1 of the supplementary material), SSIM quality constraint , and optimization iterations. Users are assigned independent 256-bit cryptographic keys . In the primary evaluation, a 32-bit user identifier is protected via a concatenated channel coding pipeline consisting of a BCH(63,36,) outer code and a repetition-5 inner code. To demonstrate high-capacity, we also evaluate a 128-bit payload mapping to 1,024 latent coordinates protected by a BCH(255,131,) and repetition-4 coding scheme.
Evaluation Metrics To evaluate visual fidelity and distributional generation quality, we assess performance using Peak Signal-to-Noise Ratio (PSNR), the Structural Similarity Index Measure (SSIM), the Fréchet Inception Distance (FID), and CLIP score. To quantify extraction robustness, we report both Bit Accuracy (the proportion of payload bits correctly decoded) and Decode Rate (the percentage of images for which the exact 32-bit user identifier is fully recovered).
We adopt decode rate as the primary metric for robustness evaluation under attacks, as it is exponentially harder than bit-level accuracy (See discussion in Section 7.1 of the supplementary material). For forgery detection, we measure the Hamming distance between the decoded perceptual hash and the target image’s recomputed hash; a distance exceeding 4 out of 16 bits indicates a detected forgery.
Performance Evaluation
Table 2 reports evaluations of image quality, decode rate, and bit accuracy across 1,500 images (500 per benchmark). Protected images are evaluated under five conditions from the WAVES benchmark (An et al. 2024): no attack, JPEG (QF=50), additive Gaussian noise (), Gaussian blur (kernel 5, ), and Cheng20 (Cheng et al. [6]) neural compression (Quality 3). We use ZoDiac’s (Zhang et al. [32]) public attack suite to guarantee absolute baseline parity. All reported results use the joint three-layer decoder for payload extraction and verification.
Image Quality Watermarked images preserve high visual fidelity (SSIM = 0.921, PSNR = 28.8 dB, FID = 8.74, CLIP score = 0.989 ± 0.011), closely aligning with the quality boundaries established by ZoDiac. Figure 3 provides qualitative examples demonstrating that the embedded perturbations remain imperceptible under standard human inspection, with structural residuals becoming apparent only under 10× error amplification.

Figure 3. Visual quality comparison across three benchmarks (one image per dataset). Left: original. Center: watermarked (SSIM ≥ 0.92). Right: absolute pixel difference amplified 10×. Rows: MS-COCO (top), DiffusionDB (middle), WikiArt (bottom).
Watermark Robustness The proposed method maintains high decode rate and bit accuracy across all four distortion categories. While Cheng20 neural compression represents the most severe degradation vector, our three-layer decoder achieves a 91.7% decode rate—a significant increase from the 81.8% baseline without Chase decoding. All other configurations exceed a 97% success rate. Chase decoding provides the largest margins of improvement under Cheng20 (+9.9%), successfully reconstructing near-threshold payloads where standard algebraic BCH decoding fails (§ 4.6).
| Dataset | Quality | Decode Rate (%) ↑ | Bit Accuracy (%) ↑ | |||||||||||
| SSIM ↑ | PSNR ↑ | FID ↓ | CLIP ↑ | None | JPEG | Noise | Blur | Ch20 | None | JPEG | Noise | Blur | Ch20 | |
| MS-COCO | 0.919 | 29.2 | 20.8 | 0.983 | 99.6 | 97.4 | 96.4 | 99.6 | 89.4 | 99.9 | 99.1 | 98.7 | 99.9 | 96.4 |
| DiffusionDB | 0.921 | 28.3 | 11.0 | 0.995 | 100 | 99.4 | 98.0 | 100 | 93.4 | 100 | 99.7 | 99.1 | 100 | 98.2 |
| WikiArt | 0.922 | 29.0 | 16.4 | 0.992 | 100 | 98.8 | 98.2 | 100 | 92.2 | 100 | 99.7 | 99.4 | 100 | 97.6 |
| Combined | 0.921 | 28.8 | 8.74 | 0.989 | 99.9 | 98.5 | 97.5 | 99.9 | 91.7 | 100 | 99.5 | 99.1 | 100 | 97.4 |
Table 2 (PDF p. 6). Evaluation of image visual quality, decode rate, and bit accuracy across three benchmarks. Image quality metrics are averaged under the baseline (no-attack) condition. “Combined” is the average across 1,500 images (500 per benchmark).
Attack Resilience
Regeneration Attack We benchmark against CtrlRegen+ (Liu et al. [13]), a state-of-the-art diffusion-based removal attack that regenerates imagery via controlled denoising on a surrogate Stable Diffusion 1.5 model. Table 3 presents robustness comparisons for regeneration across methods. At the standard quality-preserving step threshold of , our method achieves a 94.6% exact 32-bit decode rate on 500 MS-COCO images. At an aggressive setting of , the decode rate drops to 55.4%.
| Method | Setting | Metric | No Att. | s=0.3 | s=0.5 |
| Ours | Existing | Decode | 99.6% | 94.6% | 55.4% |
| ZoDiac | Existing | Det. | 99.8†% | 18.0% | 6.0% |
| StegaStamp | Existing | Det. | 100†% | ∼1% | ∼0% |
| TreeRing | Gen. | Det. | 99†% | ∼12% | ∼0% |
Table 3. Robustness under CtrlRegen+ regeneration attack (500 MS-COCO images). †From original paper. StegaStamp and TreeRing from Liu et al. [13].
While ZoDiac reports 98.8% detection under the earlier Zhao et al. vector (Zhao et al. [33], our evaluation demonstrates its accuracy falls to 18.0% against CtrlRegen+ (). StegaStamp and Tree-Ring are similarly neutralized. Our framework represents the only post-hoc image watermarking solution that preserves high multi-bit decoding capacity under state-of-the-art regeneration constraints.
Forgery Attack To assess resilience against watermark transplantation, we simulate an adversarial forgery protocol: the attacker performs DDIM inversion on a watermarked source image, extracts the latent coefficients at the 317 designated watermark coordinates, and implants them into the latent representation of an unrelated target image. We evaluate two content-fingerprinting modes independently within the same 317 positions. Table 4 reports modification survival rates and forgery detection for both modes.
| Modification | Pass / Total | Rate |
| Brightness (±0.1) | 500/500 | 100% |
| Gaussian blur () | 500/500 | 100% |
| Noise () | 494/500 | 98.8% |
| Noise () | 459/500 | 91.8% |
| JPEG (QF=50) | 481/500 | 96.2% |
| JPEG (QF=25) | 453/500 | 90.6% |
| All modifications | 2887/3000 | 96.2% |
Table 4. Fuzzy commitment: modification survival (500 images) and forgery detection comparison.
| Method | Forgery det. | Adv. resistant |
| pHash (16-bit) | 441/500 (88.2%) | ✗ |
| CLIP + RS fuzzy commit | 1210/1225 (98.8%) | ✓ |
Table 9.
(1) Perceptual hash (pHash): We construct 500 forgery pairs (50 source images × 10 targets), jointly encoding a 32-bit user identifier and a 16-bit perceptual hash. All 50 authentic images yield a hash Hamming distance of 0 with correct user identification. Under the forgery condition, the mean decoded Hamming distance rises to 7.4/16 (min 2, max 14), with 441/500 (88.2%) transplanted pairs exceeding the detection threshold. The undetected cases correspond to source–target pairs with high visual similarity, producing near-colliding perceptual hashes.
(2) Cryptographic fuzzy commitment: We evaluate CLIP-based forgery detection across 1,225 cross-image pairs (50 images, all pairwise combinations). The fuzzy commitment detects 1,210/1,225 (98.8%) forgeries, systematically addressing the hash-collision vulnerability of pHash.
The CLIP-based fuzzy commitment (§ 5 of the supplementary material) is evaluated on 500 images across six standard modifications. It achieves a 96.2% overall modification survival rate and increases forgery detection from 88.2% (pHash baseline) to 98.8% (1,210/1,225 cross-image pairs). Notably, the targeted brightness attack that neutralizes pHash at SSIM > 0.94 fails entirely against our CLIP-based features: across 500 test images, the strongest adversarial optimization only reduces the cross-image CLIP distance to 48/128, far exceeding the Reed–Solomon correction threshold. While a ∼9% false-rejection rate occurs under heavy JPEG compression (QF=25), this trade-off effectively eliminates the structural adversarial vulnerability of traditional perceptual hashing.
Model generalizability
Table 5 (PDF p. 7) shows the performance of our training-free, plug-and-play method, which can be transferred directly to other LDM variants. We test generalization to SD 1.5 (same UNet architecture, different weights). SD 1.5 achieves comparable results on 500 MS-COCO images (SSIM = 0.921, decode rates 89.8–99.6% across all attacks). The method is compatible with any LDM sharing the same VAE architecture.
| Model | SSIM | PSNR | FID | CLIP | Decode Rate (%) |
| SD 2.1 | 0.919 | 29.2 | 20.8 | 0.983 | 89.4–99.6 |
| SD 1.5 | 0.921 | 29.8 | 19.2 | 0.964 | 89.8–99.6 |
Table 5 (PDF p. 7). Performance of our method across different Stable Diffusion variants (500 MS-COCO images).
Capacity Analysis
Table 6 demonstrates the capacity–robustness trade-off by evaluating 32-bit and 128-bit user identifiers on 500 MS-COCO images. The 32-bit configuration maximizes resilience by using a repetition-5 inner code and BCH outer correction to withstand all standard distortions. The 128-bit setup offers the payload capacity—enabling simultaneous encoding of a user ID, timestamp, and metadata—by utilizing 1,024 latent coordinates (6.3% of the latent space). Remarkably, despite this larger payload footprint, robustness and visual fidelity remain nearly indistinguishable from the baseline. Both variations employ the same embedding pipeline, differing only in error-correcting code parameters.
| Config | SSIM | PSNR | None | JPEG | Noise | Blur | Ch20 |
| 32-bit (317) | .919 | 29.2 | 99.6 | 97.4 | 96.4 | 99.6 | 89.4 |
| 128-bit (1024) | .920 | 28.4 | 99.2 | 98.0 | 97.0 | 99.4 | 92.2 |
Table 6. Capacity–robustness tradeoff.
Ablation Studies
Decoder Design Impact: BCH vs. Chase Table 7 shows the impact of Chase decoding on Cheng20, the hardest attack. Without Chase, BCH alone achieves 81.8% by correcting up to 5 coded-bit errors after majority voting. Adding Chase decoding improves this to 91.7% by trial-flipping unreliable coded bits before BCH correction, rescuing images with 6–12 coded-bit errors at negligible computational cost.
| Decoder | COCO | DiffDB | WikiArt | All |
| BCH(), no Chase | 74.8 | 86.2 | 84.4 | 81.8 |
| BCH + Chase() | 89.4 | 93.4 | 92.2 | 91.7 |
Table 7. Decoder design impact on Cheng20 decode rate (hardest attack). All results on 1,500 images with the same embedded watermarks.
Comparison with Prior Work
Table 8 benchmarks our method against ZoDiac, PRC, Xiong et al., and OptMark across quality and robustness. Cross-method comparisons require care due to metric heterogeneity: ZoDiac reports zero-bit detection, PRC evaluates TPR at a fixed FPR, Xiong et al. and OptMark measure bit accuracy, and we report exact 32-bit decode rates. Exact payload decoding is strictly more challenging than detection or bit-level accuracy. Despite optimizing for this rigorous metric, our framework remains highly competitive. Our Cheng20 decode rate (91.7%) is numerically lower than ZoDiac’s 98.6%, but ZoDiac measures the presence of zero bits—a simpler task that requires only ∼60% sign alignment rather than full 32-bit recovery (Section 7 of the supplementary material). Among multi-bit baselines, our performance matches or exceeds both Xiong et al. and OptMark across all attacks (Table 8), while representing the only training-free solution applicable to existing images. PRC occupies a distinct generation-only paradigm with bounds on cryptographic undetectability, making it less directly comparable. Lastly, our lower PSNR (28.8 dB vs. 32.5 dB for Xiong et al.) represents the inherent cost of post-hoc asset modification, and our high SSIM (0.921) demonstrates that this geometric variance is perceptually minor.
| Ours | ZoDiac | PRC | Xiong | OptMark | |
| SSIM | 0.921 | 0.92 | — | 0.93 | — |
| PSNR | 28.8 | 29.4 | — | 32.5 | — |
| Decode / Bit acc. (ours) vs. reported metric | |||||
| JPEG | 98.5/99.5 | 99.2 | 1.00 | 99.4 | 99.3 |
| G-noise | 97.5/99.1 | 99.6 | 0.94 | 98.0 | — |
| G-blur | 99.9/100 | 99.6 | — | 99.6 | — |
| Ch20/VAE | 91.7/97.4 | 98.6 | — | — | 90.4 |
| Metric | Dec./Bit | Det. | TPR | Bit acc. | Bit acc. |
Table 8. Comparison with prior watermarking methods.
Conclusion
In this paper, we presented a novel watermarking framework that embeds multi-bit payloads into the latent space of arbitrary, pre-existing images without requiring any fine-tuning or retraining of the underlying generative architecture. We systematically demonstrated that the structural overcompleteness of diffusion latent spaces provides robust multi-bit watermarking under both benign downstream transformations and advanced adversarial attacks. Comprehensive experiments demonstrate that our training-free framework provides a practical, provenance-by-design defense infrastructure for enforcing content authenticity and combating digital misattribution in the generative AI era.
References
- [1]Al-Haj, A. 2007. Combined DWT-DCT digital image watermarking. Journal of computer science, 3(9): 740–746.DOI
- [3]Bistrón, M.; Žurada, J. M.; and Piotrowski, Z. 2026. Deep Learning for Image Watermarking: A Comprehensive Review and Analysis of Techniques, Challenges, and Applications. Sensors, 26(2): 444.DOI
- [4]Cao, J.; Li, Q.; Zhang, Z.; Ni, J.; and Lu, R. 2025. Secure and robust watermarking for ai-generated images: A comprehensive survey. arXiv preprint arXiv:2510.02384.arxiv.org/abs/2510.02384
- [5]Chase, D. 1972. A Class of Algorithms for Decoding Block Codes with Channel Measurement Information. volume 18, 170–182.
- [6]Cheng, Z.; Sun, H.; Takeuchi, M.; and Katto, J. 2020. Learned Image Compression with Discretized Gaussian Mixture Likelihoods and Attention Modules. In CVPR.arxiv.org/abs/2001.01568
- [8]Gunn, S.; Zhao, X.; and Song, D. 2025. An Undetectable Watermark for Generative Image Models. In ICLR.arxiv.org/abs/2410.07369
- [9]Jain, A.; Kobayashi, Y.; Murata, N.; Takida, Y.; Shibuya, T.; Mitsufuji, Y.; Cohen, N.; Memon, N.; and Togelius, J. 2025. Forging and Removing Latent-Noise Diffusion Watermarks Using a Single Image. arXiv preprint arXiv:2504.20111.arxiv.org/abs/2504.20111
- [10]Jia, Z.; Fang, H.; and Zhang, W. 2021. MBRS: Enhancing Robustness of DNN-Based Watermarking by Mini-Batch of Real and Simulated JPEG Compression. In ACM MM.DOI
- [11]Juels, A.; and Wattenberg, M. 1999. A Fuzzy Commitment Scheme. In Proceedings of the 6th ACM Conference on Computer and Communications Security (CCS), 28–36.DOI
- [12]Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Dollár, P.; and Zitnick, C. L. 2014. Microsoft COCO: Common Objects in Context. In ECCV.arxiv.org/abs/1405.0312
- [13]Liu, Y.; Song, Y.; Ci, H.; Zhang, Y.; Wang, H.; Shou, M. Z.; and Bu, Y. 2025. Image Watermarks are Removable Using Controllable Regeneration from Clean Noise. In ICLR.arxiv.org/abs/2410.05470
- [14]Müller, A.; Lukovnikov, D.; Thietke, J.; Fischer, A.; and Quiring, E. 2025. Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models. In CVPR.arxiv.org/abs/2412.03283
- [16]Nguyen-Le, H.-H.; Tran, V.-T.; Nguyen, T.; and Le-Khac, N.-A. 2025. A survey on proactive deepfake defense: Disruption and watermarking. ACM Computing Surveys, 58(5): 1–37.
- [17]Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J. K.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models from Natural Language Supervision. In ICML.arxiv.org/abs/2103.00020
- [18]Rebuffi, S.-A.; Tran, T.; Lacatusu, V.; Fernandez, P.; Souček, T.; Jovanović, N.; Sander, T.; Elsahar, H.; and Mourachko,
- [20]Song, J.; Meng, C.; and Ermon, S. 2021. Denoising Diffusion Implicit Models. In ICLR.arxiv.org/abs/2010.02502
- [21]Struppek, L.; Hintersdorf, D.; Neider, D.; and Kersting, K. 2022. Learning to Break Deep Perceptual Hashing: The Use Case NeuralHash. In ACM Conference on Fairness, Accountability, and Transparency (FAccT), 58–69.DOI
- [22]Sun, Y.; Ni, R.; Tan, C.; Liu, H.; Ni, W.; Tao, R.; and Zhao, Y. 2026. RAIN: Redundancy-Aware Latent Injection for Quality-Preserving Image Watermarking. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 35784–35792.DOI
- [23]Tancik, M.; Mildenhall, B.; and Ng, R. 2020. StegaStamp: Invisible Hyperlinks in Physical Photographs. In CVPR.arxiv.org/abs/1904.05343
- [24]Wang, Z. J.; Montoya, E.; Munechika, D.; Yang, H.; Hoover, B.; and Chau, D. H. 2023. DiffusionDB: A Large-Scale Prompt Gallery Dataset for Text-to-Image Generative Models. In ACL.arxiv.org/abs/2210.14896
- [25]Wen, Y.; Kirchenbauer, J.; Geiping, J.; and Goldstein, T. 2023. Tree-Rings Watermarks: Invisible Fingerprints for Diffusion Images. In NeurIPS.DOI
- [26]WikiArt. 2023. Visual Art Encyclopedia. https://www.wikiart.org/.
- [27]Xing, J.; Ci, H.; Xu, H.; Yuan, H.; Liu, Y.; and Shou, M. Z. 2025. OptMark: Robust Multi-Bit Diffusion Watermarking via Inference Time Optimization. arXiv preprint arXiv:2508.21277.
- [28]Xiong, C.; Qin, C.; Feng, G.; and Zhang, X. 2023. Flexible and Secure Watermarking for Latent Diffusion Model. In ACM MM.DOI
- [29]Yang, Z.; Zeng, K.; Chen, K.; Fang, H.; Zhang, W.; and Yu, N. 2024. Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models. arXiv preprint arXiv:2404.04956.arxiv.org/abs/2404.04956
- [31]Zhang, K. A.; Xu, L.; Cuesta-Infante, A.; and Veeramachaneni, K. 2019. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285.arxiv.org/abs/1909.01285
- [32]Zhang, L.; Liu, X.; Martin, A. V.; Bearfield, C. X.; Brun, Y.; and Guan, H. 2024. Attack-Resilient Image Watermarking Using Stable Diffusion. In NeurIPS.arxiv.org/abs/2401.04247
- [33]Zhao, X.; Ananth, P.; Li, L.; and Wang, Y.-X. 2023. Invisible Image Watermarks are Provably Removable Using Generative AI. arXiv preprint arXiv:2306.01953.arxiv.org/abs/2306.01953
- [34]Zhao, X.; Gunn, S.; Christ, M.; Fairoze, J.; Fabrega, A.; Carlini, N.; Garg, S.; Hong, S.; Nasr, M.; Tramèr, F.; Jha, S.; Li, L.; Wang, Y.-X.; and Song, D. 2025. SoK: Watermarking for AI-Generated Content. In IEEE Symposium on Security and Privacy (S&P). Zhou, T.; Ding, R.; Liu, G.; Fleming, C.; Kompella, R. R.; Fei, Y.; Xu, X.; and Ren, S. 2026. MetaSeal: Defending Against Image Attribution Forgery Through Content-Dependent Cryptographic Watermarks. Transactions on Machine Learning Research.
- [35]Zhu, J.; Kaplan, R.; Johnson, J.; and Fei-Fei, L. 2018. HiD-DeN: Hiding Data with Deep Networks. In ECCV.
