SAMP: An Adaptive Margin-Prioritized Framework for Robust Diffusion Watermarking
SAMP: Self-Adaptive Margin Prioritization For Robust Diffusion Watermarking
The authors propose a diffusion watermarking framework, SAMP, which adaptively adjusts embedding strength and dimension priority; they report an average M-ACC of 0.993 and a 47.1% reduction in LPIPS compared with GROW.
The defender embeds a watermark during the diffusion model's denoising process for source tracing and authenticity verification; robustness is evaluated against attacks such as image transformations (JPEG, noise, blur, crop, rotation, brightness) and diffusion-based regeneration, and the paper does not explicitly characterize the attacker's knowledge and capabilities.
- Diffusion models make forged images easier to generate and source tracing difficult. Existing watermarking methods use fixed embedding strength or treat all dimensions equally, which the authors argue introduces unnecessary distortion and provides insufficient reinforcement for vulnerable dimensions.
- SAMP embeds the watermark in the later denoising stages: global marginal adjustment determines the embedding strength according to the interval between the target codeword and competing codewords and an attack-proxy perturbation, and vulnerability-sensitive allocation adjusts the optimization weights according to each encoded bit's sensitivity to local perturbations.
- On Stable Diffusion 2.1 and two datasets, SAMP averages M-ACC 0.993 and PSNR 31.98 dB; the authors state LPIPS is reduced by 47.1% compared with GROW; extraction requires only a single VAE encoding, 0.050 s/image.
- The paper does not specifically discuss limitations. Experiments are conducted only on Stable Diffusion 2.1, with 5,000 prompts per dataset, and the number of repetitions and statistical tests are not reported.
- Evaluated models
- Stable Diffusion 2.1
- Benchmarks
- Stable-Diffusion-PromptsMS-COCO
- Metrics
- M-ACCPSNRSSIMLPIPSFID
SAMP is an adaptive framework for robust watermarking of diffusion models, comprising two mechanisms: global margin adjustment and vulnerability-sensitive allocation. The former adjusts embedding strength based on the separation between the target codeword and competing codewords and the estimated impact of common attacks, while the latter gives higher optimization priority to watermark dimensions that are easily weakened by local perturbations. Experiments show that SAMP achieves an average information recovery accuracy of 0.993, with a 16.8% improvement in PSNR and a 47.1% reduction in LPIPS over the state-of-the-art method.
Paper analysis
What problem does this paper attempt to solve?
Addressing the problems of fixed embedding strength and equal treatment of dimensions in diffusion watermarking, the authors propose an adaptive framework, SAMP.
Core problem: existing diffusion watermarks use fixed embedding strength during denoising and treat all watermark dimensions equally, making it difficult to balance distortion and robustness.
- Scenario and motivation: Diffusion models lower the barrier to forging visual content, and the authors argue that digital watermarks can directly associate source information, supporting source tracing and authenticity verification.
- Shortcomings of existing methods: The authors state that existing methods use a fixed embedding strength throughout the denoising process, or treat all watermark dimensions equally; when some dimensions are already reliable this introduces unnecessary distortion, while dimensions more vulnerable to subsequent perturbations are insufficiently reinforced.
- Core observation: The authors argue that robust watermarking requires adaptively adjusting both the embedding strength during denoising and the optimization priority across watermark dimensions.
- Proposed approach: The SAMP framework is proposed, comprising global marginal adjustment (adjusting embedding strength according to the interval between the target codeword and competing codewords and an attack-proxy perturbation estimate) and vulnerability-sensitive allocation (giving higher optimization priority to dimensions easily weakened by local perturbations).
- Threat model:
- Defender's goal: embed a watermark during diffusion model generation so that a secret message can later be recovered from the image.
- Evaluated attacks: image transformations (JPEG compression, Gaussian noise, blur, crop, rotation, brightness adjustment) and diffusion-based regeneration.
- Victim system: the Stable Diffusion 2.1 text-to-image pipeline.
- Attacker knowledge and capabilities: not explicitly characterized in the paper.
What related research exists?
The paper divides related work into two categories, initial latent state embedding and denoising trajectory embedding, and uses five methods as baselines.
Initial latent state embedding
- Tree Ring [1] (2023) — introduces structured Fourier patterns into the initial noise for watermark verification.
- Gaussian Shading [2] (2024) — encodes the watermark into the latent representation while preserving the original latent distribution; RingID [3] (2024) extends this to multi-channel identification.
- WIND [4] and PRC [5] (both ICLR 2025) — explore latent-space embedding strategies to improve robustness and generation quality.
Denoising trajectory embedding
- LFLDW [6] (ICASSP 2025) — injects the watermark signal into selected latent frequency components at intermediate denoising steps.
- GROW [7] (CVPR 2026) — progressively guides the predicted clean latent state across multiple denoising steps.
- The paper points out that such trajectory methods still rely on predefined embedding strategies, cannot adjust the intervention strength according to the current watermark reliability, and cannot prioritize more vulnerable dimensions.
Baselines and benchmarks: The baselines are five methods: Tree-Ring, Gaussian Shading, RingID, WIND, and GROW; the evaluation datasets are Stable-Diffusion-Prompts and MS-COCO (5,000 texts randomly selected from each).
The authors position SAMP as an improvement over trajectory-based methods: it dynamically controls embedding during denoising according to codeword intervals and dimension sensitivity, rather than using a predefined strategy.
How does the paper address this problem?
SAMP adjusts embedding strength at codeword intervals in the later denoising stages and allocates optimization weights according to dimension sensitivity, enabling decoding without inversion.
SAMP adaptively embeds the watermark during the diffusion denoising process: it maps the predicted clean latent state into a learned watermark coordinate space, with strength determined by the current embedding margin and direction weights determined by the bit sensitivity measured from local perturbations, and the updated latent state is sent back to the sampler.
Watermark representation and coordinate mapping
- The secret message m (L=16 bit) is mapped by an error-correcting encoder to a bipolar codeword b (Lc=32, RM(2,5)); the columns of the key-dependent watermark basis W are orthogonalized and fixed.
- A nonlinear coordinate mapping hϕ is trained offline: the composite objective Loff = Lcode + λpair·Lpair + λc·Lprobe + λpers·Lpers constrains codeword recovery, cross-carrier consistency, carrier information suppression, and coordinate stability under perturbation, respectively; training includes crop and scale perturbations, and hϕ is frozen after training.
- The embedding margin of the i-th bit is qi(z)=bi·ui, where a positive sign indicates the target sign is correct.
Global margin regulation
- At each step, the gap Gcode,t between the target codeword and the strongest competing codeword is computed, and its minimum under lightweight proxy perturbations, Gproxy,t, is taken, yielding the conservative margin Gt and the robustness gap Rt=[gsafe−Gt]+.
- The modification cost Ct is estimated and normalized using a small-direction displacement plus radius-preserving projection; the step strength is ηt = ηteff·σ(λQ·Rt − C̄t), i.e., strengthened when the margin is insufficient and suppressed when the modification cost is high.
Vulnerability-sensitive allocation
- The watermark basis is perturbed to Wk = W + UAk (U orthogonal to W), with K=2 perturbations, probed once every 5 steps; the sensitivity of each bit vi,t is the average margin difference before and after perturbation, with larger values indicating greater vulnerability.
- The weights are updated online via a state variable: wt = Clip(softmax(Bθt+1)); the gradient of the weighted loss gives the normalized write direction, which is then scaled by ηt and passed through a radius-preserving projection to obtain the updated clean latent state.
Embedding settings and decoding
- Embedding is performed in the last Te=25 steps of 50-step DDIM; γt rises linearly from 0.10 to 1.00, temperature τ=1; hϕ is pretrained on 2,000 cached latent states (Adam 5,000 steps, lr=1e-4), then fine-tuned with crop-scale for 3,000 steps (lr=2e-5).
- Decoding requires no diffusion inversion: the received image undergoes one VAE encoding, is mapped to coordinates by the frozen hϕ, and the valid codeword with the highest matching score is taken as the decoding result.
What experiments did the paper conduct?
On SD 2.1 and two datasets, compared with five baselines, SAMP averages M-ACC 0.993 and PSNR 31.98 dB, and overhead and ablation studies were also conducted.
Experimental setup
- Subject under test: Stable Diffusion 2.1, 512×512 resolution, 50-step DDIM, guidance scale 7.5.
- Datasets: Stable-Diffusion-Prompts and MS-COCO, with 5,000 texts randomly selected from each, unmodified.
- Baselines: Tree-Ring, Gaussian Shading, RingID, WIND, GROW, under the same experimental setup, with the same prompts and initial latent states.
- Metrics: Imperceptibility using PSNR, SSIM, LPIPS, FID (clean images); robustness using full-message accuracy M-ACC (the proportion of images from which the message is successfully recovered).
- Attack conditions: Clean, JPEG, Noise, Blur, Crop, Rotation, Bright, Adversarial (referred to in the main text as diffusion-based regeneration), eight in total.
- Number of repetitions and statistical tests: not reported in the paper.
Main results
According to Table 1 (Stable-Diffusion-Prompts part; M-ACC is a proportion value):
Swipe horizontally to read the table
Method PSNR↑ LPIPS↓ FID↓ Clean Crop Rotation Adversarial Avg. M-ACC Tree-Ring 15.96 0.473 25.69 0.98 0.06 0.03 0.72 0.576 Gaussian Shading 23.68 0.093 24.62 1.00 0.92 0.43 0.82 0.883 RingID 13.20 0.533 25.95 1.00 0.23 1.00 0.81 0.879 WIND 14.71 0.471 22.06 1.00 0.77 0.76 0.88 0.915 GROW 27.37 0.087 14.59 1.00 1.00 0.97 0.92 0.963 SAMP 31.98 0.046 9.52 1.00 1.00 0.98 0.97 0.993 - The trend is consistent on MS-COCO: SAMP has an average M-ACC of 0.990, PSNR 32.18 dB, FID 8.25, and GROW is 0.980 (lower half of Table 1).
- The authors state that compared with GROW, SAMP improves PSNR by 4.61 dB (16.8%) and reduces LPIPS by 47.1%, and under regeneration attack its M-ACC is 5.4% higher than the strongest baseline.
- Exceptions: on MS-COCO Rotation, SAMP is 0.96, lower than GROW and RingID at 1.00; on Stable-Diffusion-Prompts Rotation, RingID is 1.00, higher than SAMP's 0.98.
Computational cost
- Table 2: SAMP generation is 4.35 s/image (watermark overhead 0.21 s), extraction is 0.050 s/image; GROW extraction is 0.079 s, and inversion-based methods such as Tree-Ring take about 2.4–2.5 s to extract.
- The authors state that SAMP decodes directly via a single VAE encoding and a frozen coordinate network, avoiding iterative diffusion inversion.
Ablation
- Table 3: fixed embedding PSNR 23.40 dB, M-ACC 0.962; global marginal adjustment only 32.24 dB, 0.911; vulnerability allocation only 23.32 dB, 0.984; full SAMP 31.98 dB, 0.993.
- Authors' interpretation: global marginal adjustment reduces intervention when the margin is sufficient, thereby improving PSNR, but does not strengthen the vulnerability constraint, causing M-ACC to drop to 0.911; vulnerability allocation alone raises M-ACC to 0.984; combining the two balances both.
Qualitative comparison
- Figure 2 provides a qualitative comparison of generated images and residual maps (the prompt is a macro photo of a bee); the authors state that SAMP maintains visual consistency with no perceptible watermark traces. Only method names and a two-row image layout are visible in the figure, so the specific differences cannot be verified.
What are some points that can be further explored?
The paper does not specifically discuss limitations; the experiments cover only SD 2.1, two datasets, and eight attack conditions.
Limitations and future directions noted by the authors
- The paper does not specifically discuss limitations, and has no Limitations or Future Work section; the Conclusion only summarizes the method and results.
Scope of experimental coverage
- The only model tested is Stable Diffusion 2.1, at 512×512 resolution with 50-step DDIM (Section 3.1).
- The datasets are Stable-Diffusion-Prompts and MS-COCO, with 5,000 texts each (Section 3.1).
- The robustness evaluation covers eight conditions: Clean, JPEG, Noise, Blur, Crop, Rotation, Bright, Adversarial (Table 1); the paper does not state the specific attack parameters (such as JPEG quality or crop ratio).
- The embedded payload is an RM(2,5) codeword with L=16 bit and Lc=32, embedded only in the last 25 steps (Section 3.1).
- The paper does not report the number of repetitions, variance, or statistical significance tests, nor does it report different keys or multi-message scenarios.
Summarize the main content of the paper
SAMP adaptively adjusts the strength and dimension priority of diffusion watermarks; the authors report an average M-ACC of 0.993 and extraction at 0.050 s/image.
SAMP is a diffusion image watermarking framework, aimed at source tracing and authenticity verification of generated content.
- Problem: The authors argue that existing methods use fixed embedding strength and treat all watermark dimensions equally, introducing unnecessary distortion on reliable dimensions while insufficiently reinforcing vulnerable dimensions.
- Method: In the last 25 denoising steps, global marginal adjustment determines the embedding strength at each step according to the interval between the target codeword and competing codewords (including a conservative estimate under attack-proxy perturbations) and the modification cost; vulnerability-sensitive allocation measures each encoded bit's sensitivity using orthogonal-direction perturbations and adjusts the optimization weights online; decoding requires only one VAE encoding plus a frozen coordinate mapping, with no diffusion inversion.
- Results: On Stable Diffusion 2.1 and Stable-Diffusion-Prompts, SAMP achieves an average M-ACC of 0.993, PSNR 31.98 dB, LPIPS 0.046, and FID 9.52 (Table 1); the authors state that compared with GROW, PSNR improves by 16.8% and LPIPS decreases by 47.1%, and under regeneration attack M-ACC is 5.4% higher than the strongest baseline.
- Exception: On MS-COCO Rotation, SAMP (0.96) is lower than GROW and RingID (1.00).
- Cost: Generation is 4.35 s/image (watermark overhead 0.21 s), extraction is 0.050 s/image, lower than the approximately 2.4 s of inversion-based baselines (Table 2).
- Ablation: With global marginal adjustment only, M-ACC drops to 0.911; with vulnerability allocation only, PSNR drops to 23.32 dB; the authors state that the two components are complementary (Table 3).
Authors' conclusion: SAMP achieves reliable watermark recovery while maintaining image quality. The paper does not discuss limitations, and the experiments cover only a single model and a 16 bit payload.
Related papers
Show 3 more papers
- GeoDetect: Geometric Adversarial Detection for VLPsDefensesarXiv