betaFeedback

WorksComputer Science

Sequential Semantic-Non-Semantic Watermarking: Robustness Against UnMarker Attacks

Abstract

Advancements in artificial intelligence (AI) have accelerated digital content generation and intensified concerns around fraud, misuse, counterfeiting, and content provenance. Defensive image watermarking provides a practical mechanism for identifying and tracing digital content, but recent watermark-removal attacks challenge its robustness. This study investigates Sequential Semantic Non-Semantic Watermarking (SSNSW), which combines HiDDeN and StegaStamp as complementary non-semantic and semantic watermarking schemes. Using 50 images from MS-COCO and CelebA-HQ, both embedding orders are evaluated, followed by UnMarker's non-semantic and semantic attacks applied individually and sequentially. Results show that both watermarks remain highly decodable after dual embedding, with only limited embedding-order effects. Under individual attacks, one watermark generally remains strongly decodable: StegaStamp under the non-semantic attack and HiDDeN under the semantic attack. Sequential attacks are therefore required to substantially reduce both watermark signals, while attack order has a limited and configuration-dependent effect. The semantic attack also shows a broader effect across the two watermark types than the non-semantic attack. LPIPS remains relatively low in several settings, although qualitative inspection reveals noticeable distortions, particularly in selected facial images, indicating that LPIPS should not be used alone to assess visual quality. Overall, the findings show that combining complementary watermark types can increase robustness against single-stage removal attacks and motivate broader multi-layer watermarking, stronger perceptual evaluation, and further study of sequential watermark-removal strategies.

Introduction

Background

Watermarking have historically been used for embedding identification marks into physical objects [1]. Later, it became crucial for paper currencies and tackling counterfeiting[2]. During the late 1900s, digital media emerged and with it rose key concerns of pirated products [3]. Based on this necessity, the term ’digital watermarking’ was introduced by Tirkel and Osborne [4].

Digital watermarking refers to embedding invisible signatures into images that verify ownership and protect products from piracy. The difference between a visible and invisible watermark is shown in Figure 1. Digital signature marks the source of the image while preserving quality and remaining imperceptible. Digital watermarking is used for authentication, provenance tracking, fingerprinting and device control.

Illustration of visible and invisible watermarking

Figure 1. Illustration of visible and invisible watermarking. Source: adapted from [5].

The growth in technology and the emergence of generative Artificial Intelligence (genAI) have made piracy and fabrication of digital products extremely easy. The creation of AI-based content has become the core of many businesses [6]. Group-IB [7] report shows that this will continue to rise and projects that by 2027 the AI-enabled losses will reach up to $40 Billion from $12 Billion in 2023[7]. Furthermore, the global counterfeit market is projected to reach a valuation of 1.79 trillion dollars by 2030 [8]. Reports show that from the year 2022 to 2023, deepfake fraud cases increased by a staggering 1740% as scammers adopted AI tools[9].

In addition, in the first quarter of 2025, these AI-driven scams crossed a market value of $200 Million [10]. These scams also include AI-enabled video call-based impersonation and have led to heists that are worth a high amount. An example in [11] shows a $25 Million theft from a UK-based company using deepfakes. Financial Institutions report that each fraud incident creates average losses of $600 thousand dollars and multiple cases are worth over a million dollars [12] worth of losses alone. The speed and realism make these deepfake scams highly effective at extorting huge amounts of money. Deloitte reports that financial firms saw deepfake-related fraud attempts surge 700% in their contexts. [13]. Furthermore, unauthorized image reuse also directly affects creators and brands. An estimated 2.5 billion images are stolen every day, resulting in losses of approximately $600 billion [14]. AI has also made it easier to reuse brand images, causing financial losses and reducing consumer trust [15].

Due to these issues, major nations have issued regulations to make AI-generated content identifiable. The European Union

Digital watermarking market size and projected growth

Figure 2. Global digital watermarking market size and projected growth. Source: adapted from [16]

2002 Euro banknote watermark example

Figure 3. 2002 Euro banknote watermark example. Source: [2].

introduced the Artificial Intelligence Act [17] in 2023, which requires AI systems to ensure that AI-generated content is identifiable. Similarly, United States Executive Order 14110 [18] addressed the identification of AI-generated content. China’s Deep Synthesis regulations, in effect since 2023, also require AI-altered or generated media to be identifiable [19]. China extended these requirements in 2025 by requiring both visible and invisible labels for AI-generated media.

As a result, the digital watermarking market growth is expected to surge in the next few years. The digital watermarking market is valued at roughly $1.4 to $1.5 billion in 2024 [16] and is expected to reach $3 to $4 billion in size by 2030 [16] as industries invest in content tracking and anti-piracy. 63% of this revenue is due to invisible watermarking schemes that are stealthy and do not leave a mark on images [16].

Previous Watermarking methods

Traditional and Learning-Based Watermarking

Traditional watermarking methods used spatial- and frequency-domain techniques. Spatial domain approaches such as Least Significant Bit [20] were computationally inexpensive but vulnerable to compression, noise, and image transformations. Frequency-domain methods based on the Discrete Cosine Transform [21] and wavelet transforms [22] offered improved robustness. However, they could still be disrupted by filtering, compression, and geometric operations. Attackers who remove watermarks still found relatively simple ways to remove these watermarks using various noise or warping filters. Hence, interest grew in the use of Machine Learning to develop robust watermarking schemes.

These limitations motivated the adoption of deep-learning-based watermarking, beginning with convolutional approaches such as Hiding Images in Plain Sight [23]. This was an early convolutional neural network (CNN) based approach to embed one image (watermark) inside another image (original). However, the embeddings was vulnerable to removal with JPEG compression. Authors of [24] built on the idea of HIPS to introduce HiDDeN, an end-to-end trainable model that embedded a message within any image while preserving perceptual similarity to original image. It consists of an encoder that embeds the message, a noise layer during training that simulated real-world distortions like cropping, blurring and compression, and a decoder that recovers the message. Additionally, an adversarial discriminator was used to encourage encoder to produced images that looked visually similar with embedded watermarks.

The paper showed that CNNs were capable of producing robust and invisible watermarks that were not easily removed with watermarking removal attacks. Building on the HiDDeN framework, ROMark[25] used adversarial techniques to improve robustness against watermark removal attack, enhancing results. Similarly, RedMark[26] introduced a differentiable attack simulation layer to improve robustness against distortions (e.g., cropping, compression). This further improved the robustness of HiDDeN to cropping and compression attacks. Robust Invisible Video Watermarking GAN (RivaGAN) [27] introduced deep-learning-based video watermarking using attention mechanisms during adversarial training. RivaGAN produced low-distortion watermarks that resisted video processing operations.

Physical and Generative Watermarking

HiDDeN was designed to provide robustness against digital transformations and preserve image identity but, printing images and recapturing images rendered HiDDeN less effective.. They add further distortions like lighting changes, angles and required a more sophisticated noise layer. To address this issue, Tancik et al introduced StegaStamp [28], a watermarking scheme for watermarking robustness in images after they are printed out and recaptured. The encoder-decoder concept of HiDDeN is used with a more comprehensive CNN architecture. Furthermore, an approximate differentiable noise layer is used to simulate lighting, camera focus focus, colour and other recapturing distortions. The study showed that upto 56 bit message could be embedded and then decoded using smartphone cameras. StegaStamp extended deep-learning-based watermarking robustness from digital to physical print scenarios.

In recent years, generative AI (genAI) based models have started to produce realistic images. Hence, watermarking AI-based content has become necessary to reduce misuse. Ning Yu et al proposed the artificial fingerprinting method [29], which embeds model specific fingerprints during training rather than after image production. Fingerprints are added to the training data so that generated images inherit a detectable signature that can be recovered by a corresponding decoder. The method was later extended using Responsible Disclosure of Generative Models [30], making fingerprinting more scalable and providing a path to fingerprinted AI digital products. Afterwards, Google developed their own SynthID Image [31] and SynthID Text [32] that supports watermarking at an large industrial scale. They used imperceptible embeddings that enable the decoder to flag-AI generated content after image transformations. The emergence of SynthID enables deployment of genAI systems in the industry.

Retraining LLMs can be expensive and so Lukas et al, introduced Pivotal Tuning Watermarks (PTW) [33] to trains a small subset of model parameters so that generated outputs contain the embeddings. Existing image generates used PTW to train efficiently. The authors of Stable Signature[34] extended the idea to diffusion models. Stable Signature fine-tuned the decoder to producde embeddings while keeping the original architecture intact. A seperate watermark decoder is used to recover the embedding from the image which ensures that no post-processing is required. Stable Signature is highly robust and showed that it can detect watermarks with over 90% accuracy even after 90% of the image is cropped. Finally, Tree-Ring Watermarks (TRW) [35] adds the embedding in the initial noise vector of the diffusion model. The embedding resembles tree rings in frequency space and are passed through the denoising process of diffusion models and added to generated image.

Previous Watermark removal methods

Watermark removal matured alongside the watermarking schemes. Early attacks relied on compression, geometric transformations and filtering. However, encoder-decoder based watermarking was highly robust and so attackers also started using deep learning methods to remove watermarks. An illustration of a visual watermark removal is shown in Figure 4. Dekel et al. [37] used multiple images with the same visible embedded watermark to be trained and then remove these watermarks. WDNet [38] split images into watermark and background components to remove them. Similarly, Split-Refine [39] used pixel localization to locate the visible watermarks and remove them. The Prior Attack study [40] uses information of the Stable Signature model to remove Stable Signature completely when given prior knowledge of the model. However, none of these methods were designed for invisible watermarks.

Example of visible watermark detection and removal

Figure 4. Example of visible watermark detection and removal. Source: [36].

For invisible watermarks, WEvade [41] randomized the decoder’s embedded message, so that decoder accuracy falls to 50%. However, the model required access to the targeted watermarking scheme. Then, Saberi et al [42] proposed an attack that introduces Gaussian noise and then passes image through a pretrained diffusion denoiser, enabling black box removal with no prior knowledge. The attack had a high success rate against non-semantic, subtle watermarks (e.g., HiDDeN). However, against semantic (low-frequency) watermarks, the method blurred parts of the image to remove these watermarks. CtrlRegen[43] uses Control Network guided regeneration to preserve original image while destroying the watermark structure. This method is extremely effective but computationally expensive.

Overall, these methods were limited by prior knowledge, model access, computational cost, image degradation, or their inability to address both semantic and non-semantic watermarks.UnMarker [44] introduced two attacks to address these limitations. The authors argue that all watermark carriers are the spectral amplitudes. They introduced the non-semantic (high-frequency) and semantic (low frequency) attack to disrupt these frequencies that carry the watermarks.

The Non-semantic attack used Direct Fourier Loss to maximize spectral differences between the original watermarked and attacked image. This results in high frequency disruption with perceptual constrain bounds to limit visual changes. Low frequency regions are visible and so thesemantic attack used learnable filters to disrupt low frequency regions as pixel manipulation will cause visual artifacts. Perceptual constraints are applied again. Results showed the UnMarker substantially reduced detection rates of both semantic and non-semantic watermarking schemes while remaining within perceptual constrained bounds.

Research Motivation and Contributions

UnMarker introduced two separate attacks: a non-semantic attack for removing high-frequency watermarks and a semantic attack for removing low-frequency watermarks. However, its evaluation considered images containing one watermark at a time.. This motivates the investigation of whether complementary watermarks types can be embedded simultaneously to counter the UnMarker attacks. Considering this, removing only one watermark is insufficient.

The objective of this study is to investigate Sequential Semantic Non-Semantic Watermarking (SSNSW) by combining a non-semantic and a semantic watermark within the same image. The study addresses the following research questions: (1) Does the watermark embedding order influence the robustness of the SSNSW scheme? (2) Does the second embedded watermark affect the decodability of the first? (3) Does the order of the semantic and non-semantic UnMarker attacks affect watermark-removal performance?

To answer these questions, the experiments are organized into three phases. Phase 1 establishes the baseline performance of HiDDeN and StegaStamp under the corresponding non-semantic and semantic UnMarker attacks. Phase 2 evaluates dual watermarking in both embedding orders: HiDDeN followed by StegaStamp and StegaStamp followed by HiDDeN. Phase 3 applies both UnMarker attacks sequentially to the dual-watermarked images, resulting in four experiments: (1) HiDDeN →\to StegaStamp →\to semantic attack →\to non-semantic attack; (2) HiDDeN →\to StegaStamp →\to non-semantic attack →\to semantic attack; (3) StegaStamp →\to HiDDeN →\to semantic attack →\to non-semantic attack; and (4) StegaStamp →\to HiDDeN →\to non-semantic attack →\to semantic attack.

HiDDeN and StegaStamp are used as the non-semantic and semantic watermarking schemes, respectively. After each embedding and attack, the bit accuracy of both watermarks are measured to emsire they remain detectable. LPIPS is also recorded to evaluate perceptual cost of each watermarking and removal scheme. This three-phase design enables the study of baseline attack effectiveness, cross-watermark interference, embedding-order robustness, attack-order effects, and whether sequential attacks are required to disrupt both watermarks.

To the best of our knowloedge, bo study has systematically explored the combination of high- and low-frequency watermarks within a single image with sequential attacks designed to remove them. Existing works mostly evaluate watermarking schemes and removals for single-layer watermarks. Therefore, this study contributes by extending watermark removal beyond isolated watermarking and removal schemes toward multi-layered watermarking and attack piplelines.

Methodology

This section describes the watermarking and watermark-removal methods, datasets and preprocessing, evaluation metrics, and the three-phase experimental design. It also presents the implementation settings and modifications used to extend the original methods to sequential watermarking and watermark removal methods.

Selected Watermarking and Removal Methods

Recent watermarking and watermark-removal methods can be computationally intensive, making an exhaustive evaluation of all available methods outside the scope of this study. Hence, HiDDeN and StegaStamp, one state-of-the-art non-semantic and semantic watermarking schemes, respectively, are selected for the experiments in this study.

a) HiDDeN.: HiDDeN [24] pioneered the use of end-to-end deep-learning-based watermarking and serves as a building block for several later watermarking schemes. HiDDeN uses a neural network to embed its signature (binary message) in images such that the image remains imperceptible. Yet, the HiDDeN decoder can recover the signal even after it undergoes common distortions.

The overview of HiDDeN is shown in Figure 5. The architecture can be broken down to four main components: The Encoder, a Noise Layer, the Decoder, and an Adversarial Discriminator. The Encoder, EθE_\theta is tasked with encoding the message, MinM_{\mathrm{in}}, within the original image IcoI_{\mathrm{co}} to produce IenI_{\mathrm{en}}. The Noise Layer NN adds random noise mimicking real-life transformations of the watermarked image to produce InoI_{\mathrm{no}}. The Decoder DϕD_\phi retrieves the message from the distorted watermarked image InoI_{\mathrm{no}}. The Encoder and Decoder are jointly optimized through LE,D\mathcal{L}_{E,D}, which combines message-reconstruction loss, image-distortion loss, and adversarial loss. Finally, the Adversarial Discriminator AγA_\gamma distinguishes between original and encoded image LA\mathcal{L}_A, encouraging encoder to produce watermarked images that remain similar to original image. The complete HiDDeN pipeline and its training objectives are expressed compactly in 1.

Overview of the HiDDeN watermarking framework

Figure 5. Overview of the HiDDeN watermarking framework. Source: adapted from [24].

Ien=Eθ(Ico,Min),Ino=N(Ico,Ien),Mout=Dϕ(Ino),LE,D=λML∥Min−Mout∥22+λI∥Ico−Ien∥22+λGlog⁡(1−Aγ(Ien)),LA=−[log⁡(1−Aγ(Ico))+log⁡(Aγ(Ien))].(1)\begin{aligned} I_{\mathrm{en}} &= E_\theta(I_{\mathrm{co}}, M_{\mathrm{in}}), \quad I_{\mathrm{no}} = N(I_{\mathrm{co}}, I_{\mathrm{en}}), \quad M_{\mathrm{out}} = D_\phi(I_{\mathrm{no}}),\\ \mathcal{L}_{E,D} &= \frac{\lambda_M}{L}\lVert M_{\mathrm{in}}-M_{\mathrm{out}}\rVert_2^2+\lambda_I\lVert I_{\mathrm{co}}-I_{\mathrm{en}}\rVert_2^2+\lambda_G\log(1-A_\gamma(I_{\mathrm{en}})),\\ \mathcal{L}_A &= -[\log(1-A_\gamma(I_{\mathrm{co}}))+\log(A_\gamma(I_{\mathrm{en}}))]. \tag*{(1)} \end{aligned}

Following the operational taxonomy adopted by UnMarker, HiDDeN is non-semantic and hides its signature in the high-frequency regions of the image. Hence, HiDDeN is selected as the non-semantic watermarking scheme for this study.

b) StegaStamp.: StegaStamp [28] is built upon the architecture of HiDDeN with noticeable changes that make the watermarking feasible for printed images. The new functions extend the noise layer with differentiable transformations that are associated with printing, image capture and geometric distortions. Following the taxonomy adopted by UnMarker, StegaStamp is treated as a semantic watermarking scheme associated with lower-frequency watermark characteristics. StegaStamp is the only watermarking scheme that supports physical printed watermarking. Hence, due to its architectural similarity with HiDDeN and support for physical print–camera transformations, StegaStamp is selected as the semantic watermarking scheme for this study.

The working mechanism of StegaStamp is shown in Figure 6. As summarized in Eq. (2), noise simulation T\mathcal{T} is applied between the Encoder EθE_\theta and the Decoder DϕD_\phi so that the embedded message can be recovered after simulated printing and camera distortions. StegaStamp additionally uses LPIPS and geometric reconstruction losses to preserve perceptual quality and improve robustness to geometric transformations.

Working mechanism of StegaStamp

Figure 6. Working mechanism of StegaStamp, adapted from [28].

Mout=Dϕ(T(Eθ(x,Min))),(2)M_{\mathrm{out}}=D_\phi(\mathcal{T}(E_\theta(x,M_{\mathrm{in}}))), \tag*{(2)}

c) UnMarker.: The UnMarker[44] is then introduced for removing these watermarks. UnMarker is a black-box optimization algorithm that can remove semantic and non-semantic watermarks without any prior knowledge of the watermarking scheme. This makes UnMarker unique and the authors considered it the first practical attack on semantic watermarks. The goal of UnMarker is to produce an output image xunmarkx_{\mathrm{unmark}} that is perceptually close to the input watermarked image x′x' such that the decoder fails to read the hidden message mm. As discussed from LPIPS basics, here the visual difference ldl_d refers to visual difference between the watermarked image x′x' and the output image xunmarkx_{\mathrm{unmark}} and they must be less than a certain threshold tldt_{l_d}. Therefore, the objective is shown in Equation (3).

x′→xunmark,ld(x′,xunmark)≤tld,Fc(xunmark)≠Fc(x′).(3)\begin{aligned} x' &\to x_{\mathrm{unmark}},\\ l_d(x',x_{\mathrm{unmark}}) &\leq t_{l_d},\\ F_c(x_{\mathrm{unmark}}) &\neq F_c(x'). \tag*{(3)} \end{aligned}

For successful implementation, the UnMarker assumes the methods we deduced earlier in HiDDeN and StegaStamp. That is, for a watermark to be robust against distortions, they must be embedded in specific spectral amplitudes (frequency bands). They call these the universal spectral carrier. Therefore, these assumptions led to two types of watermarking schemes. The non-semantic high-frequency attack and the semantic low-frequency attacks.

Non-semantic attack. Non-semantic watermarking schemes primarily inject watermarks in the global spectral amplitudes, which are associated with high frequency characteristics. Direct Fourier Loss (DFL) measures the band-wise spectral differences between the watermarked image xwx_w and attacked image xumx_{\mathrm{um}}, which UnMarker maximizes to disrupt the watermark. To ensure visual degradation is minimal, LPIPS ℓp(⋅,⋅)\ell_p(\cdot,\cdot) constrains perceptual deviation using the threshold tpt_p, while the ℓ2\ell_2 norm threshold t2t_2 prevents the perturbation from being concentrated in only a small number of high-amplitude pixels. The attacked image xumx_{\mathrm{um}} is obtained by optimizing DFL while penalizing any violation of the LPIPS and ℓ2\ell_2 thresholds as expressed in Eq. (4). Watermarks like HiDDeN in the high-frequency bands are significantly changed, reducing the decoders ability to recover the encoded message.

xum=xw+δ∗,δ∗=arg⁡min⁡δ [cpReLU⁡(ℓp(xw+δ,xw)−tp)+c2ReLU⁡(∥xw+δ−xw∥2−t2)−DFL⁡(xw+δ,xw)].(4)\begin{aligned} x_{\mathrm{um}} &= x_w+\delta^*,\\ \delta^* &= \underset{\delta}{\arg\min}\ \bigl[c_p\operatorname{ReLU}\bigl(\ell_p(x_w+\delta,x_w)-t_p\bigr)\\ &\quad+c_2\operatorname{ReLU}\bigl(\lVert x_w+\delta-x_w\rVert_2-t_2\bigr)-\operatorname{DFL}(x_w+\delta,x_w)\bigr]. \tag*{(4)} \end{aligned}

Semantic attack. Semantic watermarking schemes like StegaStamp embed watermarks into the low-frequency components of the image. As low-frequency watermarks are woven into the image textures with smooth transitions, the pixel-by-pixel manipulation of a non-semantic attack may fail to remove the watermark or introduce large visual artifacts. Therefore, UnMarker proposes a semantic attack in which groups of neighbouring pixels are manipulated simultaneously using optimizable filters to introduce spatially smooth perturbations. As shown in Eq. (5), the filtered image is obtained by applying TT optimizable filters with learnable parameters θ\theta to the watermarked image xwx_w with additive perturbation δ\delta. Each filtered pixel is calculated as a combination of its neighbours, producing spatially smooth intensity changes across the image. This disrupts low-frequency regions and reduce the decoder’s ability to read the watermark.

DFL introduced in non-semantic attacks maximizes pixel-wise spectral changes, but it does not apply to neighbourhood-level perturbations. Hence, UnMarker introduces Filter Regularization Loss (FRL) and Mean Pool Loss (MPL). FRL compares each pixel to its neighbours and penalizes large deviations as these outliers represent visual artifacts. However, FRL is limited to local medians; when multiple pixels are perturbed together, the local median may also change, which allows the watermark to remain intact. MPL addresses this limitation by comparing patch averages within a L×P(5×5)L \times P(5 \times5) window. UnMarker maximizes MPLL×P(xw,x)\mathrm{MPL}_{L\times P}(x_w,x), forcing the output patches to deviate from those of the watermarked image. Together, MPL disrupts low-frequency patterns while FRL prevents visual artifacts and uniform brightness shifts. The objective in Eq. (5) jointly optimizes θ\theta and δ\delta, penalizes violations of the LPIPS perceptual threshold and large ℓ2\ell_2 deviations, and maximizes the MPL and FRL terms.

xθ,δ=OF^θ(T)(xw+δ),(θ∗,δ∗)=arg⁡min⁡θ,δ [cℓDVLReLU⁡(ℓDVL(xw,xθ,δ)−tℓDVL)+λ2∥xθ,δ−xw∥2−MPLL×5×5(xw,xθ,δ)−λFRLFRL(xθ,δ)],xums=OF^θ∗(T)(xw+δ∗).(5)\begin{aligned} x_{\theta,\delta} &= \widehat{\mathrm{OF}}_{\theta}^{(T)}(x_w+\delta),\\ (\theta^*,\delta^*) &= \underset{\theta,\delta}{\arg\min}\ \bigl[c_{\ell_{\mathrm{DVL}}}\operatorname{ReLU}\bigl(\ell_{\mathrm{DVL}}(x_w,x_{\theta,\delta})-t_{\ell_{\mathrm{DVL}}}\bigr)\\ &\quad+\lambda_2\lVert x_{\theta,\delta}-x_w\rVert_2-\mathrm{MPL}_{L\times5\times5}(x_w,x_{\theta,\delta})\\ &\quad-\lambda_{\mathrm{FRL}}\mathrm{FRL}(x_{\theta,\delta})\bigr],\\ x_{\mathrm{um}}^s &= \widehat{\mathrm{OF}}_{\theta^*}^{(T)}(x_w+\delta^*). \tag*{(5)} \end{aligned}

Datasets used

To evaluate the watermarking schemes and UnMarker effectively, datasets with natural scenes and structured facial image needs to be considered. Hence, for this study, two widely-used datasets, MS-COCO and CelebA-HQ, are considered. MS-COCO[45] is a high-quality dataset that contains complex real-world scenes with detailed backgrounds and diverse objects. The diversity enables the robustness of watermarking schemes to be evaluated across heterogeneous visual content. CelebA-HQ[46] consist of high-quality facial imaging containing smooth gradients and facial patterns. These characteristics allow watermark behaviour to be evaluated in both smooth low-frequency regions, such as lighting, shading, and colour, and high-frequency regions, such as hair and eyelashes. Facial images also provide a structured image domain relevant to deepfake-related applications.

For this study, 50 images are sampled. 25 from MS-COCO and 25 from CelebA-HQ. The computationally constrained evaluation scale follows the compact experiment scale of UnMarker[44], which evaluates each watermarking system on 100 randomly selected images. The images are randomly sampled using a fixed seed to ensure reproducibility. They are then resized and cropped to a fixed resolution compatible with both watermarking schemes. The same set of images is used for all watermarking and attack and defense configurations to ensure consistency and a fair comparison.

Evaluation criteria

WAVES Benchmark[47] standardized metrics for future watermark attacks. The study finds that ’rinsing’ and other techniques can remove stubborn semantic watermarks like TRW and StegaStamp for higher distortion within the image. Furthermore, WAVES mention that the success metrics of watermarking must also ensure image quality. This benchmarking is key to future watermark attack schemes.

In this study, efficiency of watermark removal is evaluated using two criteria: (1) decoding performance via bit accuracy, and (2) visual fidelity via perceptual similarity. The two methods are discussed in further detail below.

1) Bit accuracy: Defensive watermarking methods consist of embedding a binary message mm into an image xx to produce the watermarked image x′x'. The watermark perturbation Δx\Delta x can be decoded by the decoder D(⋅)D(\cdot) to recover the binary message mm. During the experiments, the decoder outputs bit probabilities pj(I)p_j(I) and uses a threshold of 0.5 to make binary decisions. Bit Accuracy (BA) measures the success rate of the decoder and is defined as

m^j(I)=I[pj(I)≥0.5],BA(I)=1L∑j=1LI[m^j(I)=mj],(6)\hat{m}_j(I) = \mathbb{I}[p_j(I) \ge0.5], \qquad\mathrm{BA}(I) = \frac{1}{L}\sum_{j=1}^{L}\mathbb{I}[\hat{m}_j(I) = m_j], \tag*{(6)}

where LL is the message length and I[⋅]\mathbb{I}[\cdot] is the indicator function. The formula returns the proportion of the message that the decoder can retrieve in the range [0,1][0,1], where 1 corresponds to 100% of the message being recovered.

In the experiments, Bit Accuracy is evaluated in three stages: BAorig=BA(x)\mathrm{BA}_{\mathrm{orig}} = \mathrm{BA}(x), BAwm=BA(x′)\mathrm{BA}_{\mathrm{wm}} = \mathrm{BA}(x'), and BAatt=BA(xatt)\mathrm{BA}_{\mathrm{att}} = \mathrm{BA}(x_{\mathrm{att}}). Firstly, BAorig\mathrm{BA}_{\mathrm{orig}} acts as the baseline sanity check. Due to decoder bias, this value is crucial for evaluating the failure of the decoder’s performance. Secondly, BAwm\mathrm{BA}_{\mathrm{wm}} shows whether the watermark is functioning appropriately and should ideally be close to 1. Finally, BAatt\mathrm{BA}_{\mathrm{att}} quantifies the effectiveness of the attack in removing the watermark and should ideally be close or equal to BAorig\mathrm{BA}_{\mathrm{orig}}.

2) LPIPS Similarity: The constraint of watermark removal is to ensure that the attacked image remains visually as close as possible to the watermarked image while message recovery is disrupted. Hence, image fidelity is evaluated using Learned Perceptual Image Patch Similarity (LPIPS). LPIPS measures the perceptual distance between two images using CNN-based feature representations. For the watermarked image x′x' and attacked output xattx_{\mathrm{att}}, LPIPS is defined in Eq. (7).

LPIPS(x′,xatt)=∑ℓ1HℓWℓ∑h,w∥wℓ⊙Δf^ℓ,hw∥22\mathrm{LPIPS}(x',x_{\mathrm{att}})=\sum_{\ell}\frac{1}{H_{\ell}W_{\ell}}\sum_{h,w}\left\lVert w_{\ell}\odot\Delta\hat{f}_{\ell,hw}\right\rVert_2^2
Δf^ℓ,hw=f^ℓ(x′)hw−f^ℓ(xatt)hw.(7)\Delta\hat{f}_{\ell,hw}=\hat{f}_{\ell}(x')_{hw}-\hat{f}_{\ell}(x_{\mathrm{att}})_{hw}. \tag*{(7)}

where f^ℓ(⋅)\hat{f}_{\ell}(\cdot) denotes the channel-normalized feature activation at layer ℓ\ell, wℓw_{\ell} contains the learned channel-wise weights, HℓH_{\ell} and WℓW_{\ell} are the spatial dimensions of the corresponding feature map. Smaller LPIPS values indicate greater perceptual similarity. UnMarker uses LPIPS as a perceptual constraint during optimization, while this study reports it as the image-fidelity metric. An attack is considered effective when BAatt\mathrm{BA}_{\mathrm{att}} approaches the clean-image baseline BAorig\mathrm{BA}_{\mathrm{orig}} while LPIPS(x′,xatt)\mathrm{LPIPS}(x',x_{\mathrm{att}}) remains low.

Experimental design

The objective of this study is to evaluate the effects of combining two watermarks, the robustness of the watermarks under UnMarker attacks, and the impact of embedding and attack order. The same benchmark of N=50N=50 images from MS-COCO and CelebA-HQ is used. The two datasets contain objects and faces to ensure that the results do not become domain-specific. Results are reported together and separately for each dataset. The experiments are divided into three phases.

a) Phase 1: Individual watermark and attack baselines.: The effectiveness of the semantic and non-semantic UnMarker attacks is first evaluated. HiDDeN is tested against the non-semantic attack, while StegaStamp is tested against the semantic attack. The phase provides the baseline results of the study.

b) Phase 2: Dual-watermark embedding and interference study.: For Phase 2, dual watermarking schemes are implemented to provide images with two layers of defensive watermarks. This experiment studies the impact of embedding sequence and evaluates the interference caused when the second watermark overlaps the first. A bird’s-eye view of this phase is shown in Figure 7.

Dual-watermark embedding pipelines used in Phase 2

Figure 7. Dual-watermark embedding pipelines used in Phase 2.

Two dual-watermarking pipelines are applied to the benchmark dataset of N=50N=50 images {xi}i=1N\{x_i\}_{i=1}^{N}. In HiDDeNStega, each image xix_i is watermarked with HiDDeN to produce xiHx_i^{H} and then with StegaStamp to produce xiHSx_i^{HS}. The resulting dataset XHSX^{HS} is referred to as the HiDDeNStega dataset. In StegaHiDDeN, each image is watermarked with StegaStamp to produce xiSx_i^{S} and then with HiDDeN to produce xiSHx_i^{SH}. The resulting dataset XSHX^{SH} is referred to as the StegaHiDDeN dataset.

After each embedding stage, the Bit Accuracy of HiDDeN and StegaStamp is evaluated alongside LPIPS. These measurements evaluate how much the second embedding affects the first watermark and whether the images remain perceptually similar after dual watermarking.

c) Phase 3: Dual-watermark removal and attack-order study.: This phase evaluates the robustness of the dual-watermarking schemes and the effectiveness of UnMarker when both types of watermarks are present. The main outcome is to determine whether both attacks are necessary and, if so, which attack sequence is more effective. A bird’s-eye view of Phase 3 is shown in Figure 8.

Sequential attack pipelines used in Phase 3

Figure 8. Sequential attack pipelines used in Phase 3.

For each dual-watermarked dataset, XHSX^{HS} and XSHX^{SH}, two sequential attack orders are applied: non-semantic followed by semantic, and semantic followed by non-semantic. This produces four attack pipelines, covering both attack orders for both dual-watermarked datasets.

For each embedding and attack stage, the Bit Accuracy of the HiDDeN and StegaStamp decoders and LPIPS are recorded. This enables analysis of the drop in decoder performance after each attack, the effectiveness of each attack sequence, whether sequential attacks are necessary, and their impact on perceptual similarity.

Results

The results are broken down into three phases based on the experimental design of the study. First, the results of the baseline watermarking schemes are provided against the basic UnMarker attacks. Then, Phase 2 shows the results of the dual watermarking schemes. Finally, Phase 2 results are used in Phase 3 to evaluate the SSNSW schemes against the sequential UnMarker attacks.

Phase 1: Baseline study

Phase 1 is further broken down into two experiments: HiDDeN watermarking scheme being tested by the non-semantic attack, followed by the semantic attack on StegaStamp.

Table I shows the results on the benchmark dataset of the HiDDeN watermarking scheme followed by the UnMarker attack.

SubsetBAorig_{\mathrm{orig}}BAwm_{\mathrm{wm}}BAatt_{\mathrm{att}}LPIPS
COCO (1–25)0.5811.0000.5490.070
CelebA (26–50)0.5841.0000.5670.082
Overall (1–50)0.5831.0000.5580.076

Table I. Results for base HiDDeN under the non-semantic attack.

attack. On the original dataset, the HiDDeN decoder bit accuracy of HiDDeN was quite low at approximately 0.58 for both datasets. Then the watermarking worked perfectly and all the images are marked at 1.00, which is fully decodable. The non-semantic UnMarker attacks reduced the decoder readability close to original baseline, while LPIPS remained low. However, some images retained traces of HiDDeN with a decoder accuracy of 0.70. The attack produced a lower average bit accuracy on COCO, while CelebA had a slightly higher average LPIPS score. Figures 9 and 10 show the images with the highest and lowest attacked bit accuracies respectively.

Original, watermarked, and removed image with highest attacked bit accuracy

Figure 9. Example with highest attacked bit accuracy (BAatt_{\mathrm{att}}) under the non-semantic attack on the HiDDeN-embedded dataset.

Original, watermarked, and removed image with lowest attacked bit accuracy

Figure 10. Example with lowest attacked bit accuracy (BAatt_{\mathrm{att}}) under the non-semantic attack on the HiDDeN-embedded dataset.

Table II displays the results of the semantic attack on the StegaStamp-embedded dataset. Similar to HiDDeN, StegaStamp ensured full decodability with perfect bit accuracy. The semantic attack reduced the effectiveness of StegaStamp, although some traces of the watermark remained. An average bit accuracy below 0.6 suggests that the majority of the watermark was removed. The attack produced a lower average bit accuracy on the CelebA dataset. Some images retained a high bit accuracy of 0.690. The differences are shown in Figure 11, with the lowest attacked bit accuracy, and Figure 12, with the highest attacked bit accuracy.

Example with lowest attacked bit accuracy under the semantic attack on the base StegaStamp dataset

Figure 11. Example with lowest attacked bit accuracy (BAattBA_{\mathrm{att}}) under the semantic attack on the base StegaStamp dataset.

Example with highest attacked bit accuracy under the semantic attack on the base StegaStamp dataset

Figure 12. Example with highest attacked bit accuracy (BAattBA_{\mathrm{att}}) under the semantic attack on the base StegaStamp dataset.

Phase 2: Watermarking Sequence Study

Phase two of the experiments test the proposed dual watermarking study. HiDDeN and StegaStamp are applied sequentially to the benchmark dataset in both sequences. HiDDeNStega represents HiDDeN embedding followed by StegaStamp, while StegaHiDDeN represents StegaStamp embedding followed by HiDDeN.

SubsetBAorigBA_{\mathrm{orig}}BAwmBA_{\mathrm{wm}}BAattBA_{\mathrm{att}}LPIPS
COCO (1–25)0.4960.9990.6020.067
CelebA (26–50)0.5201.0000.5630.072
Overall (1–50)0.5081.0000.5830.070

Table II. Results for base StegaStamp under the semantic attack.

1) Phase 2.1: HiDDeNStega: In this section, the dataset is first embedded with HiDDeN. Afterwards, StegaStamp is applied to the dataset to create the HiDDeNStega dataset. Sample result of the dual watermarked image is shown in Figure 13. Results from applying HiDDeN watermarking on the original dataset are provided in Table III. Results show that embedding HiDDeN alone is highly successful as it provides very low visual perception change while the HiDDeN watermark remains highly decodable. Results from applying StegaStamp on the HiDDeN embedded dataset to create the HiDDeNStega scheme is provided in Table III. This dataset is referred to as the HiDDeNStega dataset (HSD) going forward.

StageSubsetBAHBA_{\mathrm{H}}BASBA_{\mathrm{S}}LPIPS
After HCOCO0.9980.5030.038
After HCelebA0.9940.5230.052
After SCOCO0.9760.9820.142
After SCelebA0.9860.9810.103

Table III. Phase 2 results for the HiDDeN→StegaStamp embedding order. H and S denote HiDDeN and StegaStamp, respectively.

HiDDeNStega watermarking pipeline showing the original image, after HiDDeN embedding, and after subsequent StegaStamp embedding

Figure 13. HiDDeNStega watermarking pipeline (Phase 2): original image, after HiDDeN embedding, and after subsequent StegaStamp embedding.

The results show that, on average, the StegaStamp embeddings provide a much higher perception change than HiDDeN. This is consistent with the base experiments, where more significant distortion was produced with semantic watermarks. However, the important finding is that the decoder scores of HiDDeN still remain relatively perfect with a bit accuracy of over 97% on average. Therefore, the theory confirms that dual watermarking schemes are applicable if they are embedded in different frequency domains of the image. Figure 13 shows the three stages of the image during the HiDDeNStega watermarking scheme. The blurring effects of StegaStamp are

slightly noticeable, while the HiDDeN watermarked images seem relatively similar. Despite this, the images can still be used to prevent scammers from pirating the image.

2) Phase 2.2: StegaHiDDeN: In this section, the original dataset is first embedded with StegaStamp. Afterwards, HiDDeN is applied on the StegaStamp embedded dataset to create the StegaHiDDeN dataset. Figure 14 shows a sample result of the dual watermarking process. Results from applying StegaStamp watermarking on the original dataset is provided in Table IV. As seen in the HiDDeNStega example, the perceptual difference is comparatively higher for StegaStamp embeddings. However, the decoder perfectly reads the watermark for StegaStamp. Results from applying HiDDeN on the Stega dataset to create the StegaHiDDeN scheme are provided in Table IV. This dataset is referred to as the StegaHiDDeN dataset (SHD) going forward.

StageSubsetBAHBA_{\mathrm{H}}BASBA_{\mathrm{S}}LPIPS
After SCOCO0.4370.9980.179
After SCelebA0.4591.0000.145
After HCOCO1.0000.9940.035
After HCelebA0.9930.9920.052

Table IV. Phase 2 results for the StegaStamp→HiDDeN embedding order. H and S denote HiDDeN and StegaStamp, respectively.

Original, StegaStamp watermarked, and StegaHiDDeN watermarked images

Figure 14. StegaHiDDeN watermarking pipeline (Phase 2): original image, after StegaStamp embedding, and after subsequent HiDDeN embedding.

Once again, the theory is confirmed that sequential watermarking can be used as one watermark does not disrupt the other one. Both watermarks remain almost perfectly decodable regardless of the order they are used in.

Phase 3: Dual-watermark removal and attack-order study

a) HiDDeNStega.: Tables V and VI show the results of applying both attack sequences to the HiDDeNStega dataset. The sample results are shown in Figure 15. The non-semantic attack followed by the semantic attack reduces the bit accuracies of both watermarks. Similarly, applying the semantic attack first and the non-semantic attack afterwards also reduces StegaStamp and HiDDeN respectively, but traces of StegaStamp remain strong.

HiDDeNStega attack pipelines

Figure 15. HiDDeNStega attack pipelines (Phase 3). Top row: Non-semantic →\rightarrow Semantic. Bottom row: Semantic →\rightarrow Non-semantic.

StageSubsetBAH\mathrm{BA}_{H}BAS\mathrm{BA}_{S}LPIPS
After NSCOCO0.5840.9540.009
After NSCelebA0.6140.9600.010
After SCOCO0.5540.7200.081
After SCelebA0.5930.6570.085

Table V. Phase 3 results for HiDDeNStega under the non-semantic→\rightarrowsemantic attack order. H and S denote HiDDeN and StegaStamp, respectively.

StageSubsetBAH\mathrm{BA}_{H}BAS\mathrm{BA}_{S}LPIPS
After SCOCO0.8580.6710.087
After SCelebA0.9090.6130.080
After NSCOCO0.5910.6830.007
After NSCelebA0.6220.6400.009

Table VI. Phase 3 results for HiDDeNStega under the semantic→\rightarrownon-semantic attack order. H and S denote HiDDeN and StegaStamp, respectively.

b) StegaHiDDeN.: Results from applying both attack sequences to the StegaHiDDeN dataset are shown in Table VII and Table VIII. Sample results are shown in Figure 16. The results show that both sequence of attacks (NS followed by S and S followed by NS) can successfully reduce bit accuracies of the watermarks. However, it is to be noted that some traces of watermarks remain strongly, especially StegaStamp, the semantic watermark similar to the case observed above.

StageSubsetBAH\mathrm{BA}_{H}BASt\mathrm{BA}_{St}LPIPS
After NSCOCO0.6400.9890.009
After NSCelebA0.6770.9880.010
After SemCOCO0.6030.6190.155
After SemCelebA0.6280.6140.140

Table VII. Phase 3 results for StegaHiDDeN under the non-semantic→\rightarrowsemantic attack order. H and St denote HiDDeN and StegaStamp, respectively.

StageSubsetBAH\mathrm{BA}_{H}BASt\mathrm{BA}_{St}LPIPS
After SemCOCO0.8800.5810.144
After SemCelebA0.9140.5800.128
After NSCOCO0.6020.5940.009
After NSCelebA0.6210.5950.010

Table VIII. Phase 3 results for StegaHiDDeN under the semantic→\rightarrownon-semantic attack order. H and St denote HiDDeN and StegaStamp, respectively.

StegaHiDDeN attack pipelines showing dual-watermarked images before and after non-semantic and semantic attacks

Figure 16. StegaHiDDeN attack pipelines (Phase 3). Top row: Non-semantic → Semantic. Bottom row: Semantic → Non-semantic.

Discussion

This section discusses the performance of the individual attacks, dual-watermark embedding, sequential watermark removal, and perceptual quality. The limitations and future scope of the study are discussed afterwards.

Baseline Watermarking and Individual Attack Performance

Phase 1 shows that both watermarking schemes are consistent and provide bit accuracies of 1.0 before the attack. The corresponding non-semantic and semantic attacks substantially reduce the bit accuracy of the watermarks; HiDDeN’s bit accuracy drops from BAwm=1.000\mathrm{BA}_{\mathrm{wm}} = 1.000 to BAatt=0.558\mathrm{BA}_{\mathrm{att}} = 0.558, is close to the original-image baseline of BAorig=0.583\mathrm{BA}_{\mathrm{orig}} = 0.583. However, the per-image results show some variation, as certain images retain traces of the watermark with a bit accuracy of up to 0.70 after the attack.

Similar results are seen when StegaStamp watermarked are subjected to the semantic attack. The bit accuracy drops from 1.000 to 0.583, approaching the original-image baseline of 0.508. Even then, certain images contain traces of the watermark with a bit accuracy of up to 0.69, but still not as strong as it was before the attacks. These findings can be studied further to determine why traces of watermarks remain. These findings show that both attacks are effective overall.

WMAttackBAorigBA_{\mathrm{orig}}BAwmBA_{\mathrm{wm}}BAattBA_{\mathrm{att}}LPIPS
HNS0.5831.0000.5580.076
SSem0.5081.0000.5830.070

Table 9. Baseline results over COCO and CelebA. H and S denote HiDDeN and StegaStamp; NS and Sem denote non-semantic and semantic attacks, respectively.

Dual-Watermark Embedding and Embedding-Order Effects

The second phase of experiments tests the proposed dual-watermarking study. HiDDeN and StegaStamp are applied in both sequences. HiDDeNStega represents HiDDeN embedding followed by StegaStamp, while StegaHiDDeN represents the reverse order. Results from phase 2 in Table III and Table IV support the hypothesis that two watermarks targeting different image characteristics can be embedded together without significant disruption to each other. For both sequences, the watermarks remain decodable, and the reduction in bit accuracy of the first watermark after embedding the second one is minimal.

HiDDeNStega achieves final bit accuracies of 0.981 and 0.982 for HiDDeN and StegaStamp, respectively. Applying StegaStamp after HiDDeN slightly decreases the HiDDeN bit accuracy, but both watermarks remain almost fully decodable. The StegaHiDDeN results are more encouraging, with final bit accuracies of 0.997 and 0.993 for HiDDeN and StegaStamp, respectively. Here, HiDDeN’s low-distortion properties result in only a small reduction in StegaStamp bit accuracy. Therefore, embedding with HiDDeN last creates slightly better decodability. Embedding order has a small effect on the final decoder performance. For the pair in this experiment, StegaStamp followed by HiDDeN provides slightly better final bit accuracies for both decoders.

Embed.Attack orderStageBAHBA_{\mathrm{H}}BASBA_{\mathrm{S}}LPIPS
H→S–WM0.9810.9820.123
H→SNS→SAfter NS0.5990.9570.009
H→SNS→SAfter S0.5730.6890.083
H→SS→NSAfter S0.8830.6420.083
H→SS→NSAfter NS0.6070.6620.008
S→H–WM0.9970.9930.044
S→HNS→SAfter NS0.6590.9880.009
S→HNS→SAfter S0.6150.6170.148
S→HS→NSAfter S0.8970.5810.136
S→HS→NSAfter NS0.6110.5940.009

Table 10. Phase 3 summary over COCO and CelebA. H→S and S→H denote the HiDDeN–StegaStamp embedding orders; NS and S denote non-semantic and semantic attacks, respectively.

Sequential Watermark Removal and Attack-Order Effects

Phase 3 evaluates the dual-watermarking schemes under multiple attacks. The first question is whether they can withstand individual UnMarker attacks and still carry traces of at least one watermark. Their robustness is then tested further using sequential attacks on both the HiDDeNStega and StegaHiDDeN schemes.

The results in Table X support the general findings of UnMarker [44]. The semantic attack is highly effective against StegaStamp. Additionally it also consistently reduces the BA of HiDDeN. This is a notable finding indicating the partial effectiveness of the UnMarker semantic attack on a non-semantic watermark. Meanwhile, the non-semantic attack has only a limited effect on StegaStamp bit accuracy. This suggests that, among the two individual attacks evaluated, the semantic attack has the broader effect against dual-watermarked images.

Results from Phase 3 show that dual-watermarking schemes can be attacked by sequential UnMarker attacks successfully, substantially reducing both watermark signals as shown in Table X. However, in some images, traces of one or both watermarks remain. This should be evaluated further in future studies. The experiments show that when two types of watermarks are present, a single attack does not consistently reduce both watermarks; one remains strongly decodable. Therefore, aligning with WAVES [47], a second attack is necessary to substantially weaken both watermark signals.

Both attack orders reduce the BA of HiDDeN and StegaStamp, although their relative effectiveness varies across embedding orders and decoders. For HiDDeNStega, one attack order produces a lower final HiDDeN BA, while the reverse order produces a lower StegaStamp BA. For StegaHiDDeN, applying the semantic attack followed by the non-semantic attack performs slightly better overall. Therefore, attack order has a limited and configuration-dependent effect.

The semantic attack causes more distortion but affects both decoders, whereas the non-semantic attack primarily affects HiDDeN. The results therefore support the semantic attack as the broader individual attack. Success against dual-watermarking schemes strengthens the case for UnMarker and other adversarial optimization attack schemes that operate without prior knowledge of the watermarking schemes.

Based on the results from this study, the UnMarker attacks act as complementary frequency disruptors. UnMarker formulates them to disrupt non-semantic and semantic watermark characteristics separately. HiDDeN and StegaStamp are associated with different regions of the frequency spectrum. The SSNSW schemes ensure that neither individual UnMarker attack consistently reduces both watermarks close to their respective original-image baselines. This increases robustness against watermark removal by motivating the use of two complementary attacks for complete watermark removal.

WAVES [47] argues that quality cost must be accounted for when robustness is increased. The robustness added should be compared with the corresponding loss in visual quality. In this study, combining HiDDeN and StegaStamp increases robustness against the tested attacks because two complementary attacks are required to substantially reduce both watermarks. The Phase 2 results also show that the second watermark can be added with limited interference and perceptual cost. Therefore, this study is aligned with the WAVES quality–robustness principle.

Perceptual Quality and LPIPS

Across the experiments, LPIPS indicates limited perceptual change. However, qualitative inspection shows that it does not always reflect visually noticeable distortion.

In Phase-1, the CelebA dataset has slightly higher LPIPS than COCO, suggesting that more distortions occur in images with facial features. This is in line with the findings from UnMarker [44] that the non-semantic attack creates high-frequency disruptions that can distort watermarks like HiDDeN while visual fidelity is controlled. For StegaStamp under the semantic attack, the quality inspection shows a different finding. The facial image in Figure 11 shows considerable distortion in both the watermarked and watermark-removal images. In contrast, the object image in Figure 12 shows substantially less visible distortion despite a similar LPIPS value. These observations suggest that the relationship between LPIPS and visible distortion is content-dependent, with a larger discrepancy in the selected facial images.

The Phase 2 results further show that the perceptual difference is not symmetric between the two watermarking schemes. HiDDeN mainly introduces a blurring effect, while StegaStamp creates more noticeable changes around image structures and edges. Despite this blurring effect, the overall content of the selected image remains largely unchanged after HiDDeN embedding. StegaStamp has a much larger LPIPS, although the distortion is less noticeable in the selected COCO example. This is consistent with StegaStamp’s objective of maintaining decodability after physical transformations such as printing [28]. For the StegaHiDDeN pipeline, the CelebA example in Figure 14c shows a similar finding. Despite the relatively low LPIPS values, the impact on the facial image is noticeable. StegaStamp does create visual artifacts, while HiDDeN only creates a blurring effect. This supports the Phase 1 observation that the selected facial images show a larger discrepancy between LPIPS and qualitative visual quality.

Similar to semantic watermarking, the semantic attack also produces higher visual distortion than the non-semantic attack despite relatively low LPIPS values. LPIPS values are compared with qualitative inspection using Figure 15 and Figure 16 that represent attacks using both sequences. CelebA images are used in these examples to further examine the facial-image behaviour observed in the earlier phases. Interestingly, the semantic attack does not produce artifacts similar to the semantic watermarking scheme StegaStamp when LPIPS is low. It is observed that the comparatively high LPIPS significantly distorts the facial image. In these selected attack examples, LPIPS corresponds more closely with the visible distortion than in the watermarking examples. Nevertheless, an LPIPS value of approximately 0.15 still corresponds to considerable distortion in the selected facial image.

The attacks were configured to produce LPIPS values comparable to those introduced by the watermarking schemes to keep the balance between attack and defense. Therefore, for semantic attacks, some higher visual distortion was allowed in some cases, as StegaStamp values tend to be very high for perfect BA scores. This is in line with the objective of StegaStamp [28]. Overall, LPIPS remains useful for comparing perceptual change across the experiments, but the qualitative examples show that it should not be used alone to assess visual quality. In particular, the selected facial images show larger discrepancies between LPIPS and visually noticeable distortion than the selected object images.

Limitations, gaps and future works

The study is limited to evaluating only two watermarking schemes for SSNSW schemes due to the computational resources required for the experiments. HiDDeN and StegaStamp were selected as the non-semantic and semantic watermarking schemes, respectively. However, the success of other semantic watermarking schemes cannot be concluded from the results of StegaStamp alone. Future studies should therefore evaluate SSNSW using other watermarking methods, such as Stable Signature and Tree-Ring Watermarks. Furthermore, the attack sequence in this study is limited to UnMarker attacks. Therefore, the robustness conclusions are limited to the evaluated UnMarker setting, and future work should test SSNSW against other watermark-removal and evasion attacks.

The second limitation concerns perceptual evaluation. As discussed, LPIPS does not always correspond to visual quality in facial-image examples. LPIPS provides a consistent metric of perceptual change. However, it should not be used alone to evaluate visual fidelity. Future studies should incorporate human-grounded perceptual evaluation alongside LPIPS, aligning with WAVES benchmarking. Alternatively, methods such as R-LPIPS [48] can be investigated to determine whether they capture distortions better in face images.

The computational cost also limited the experiments to 50 images from two datasets. Each semantic attack on the 50 images required approximately three to four hours on the FAU HPC system. While the paired experimental design allows trends to be observed within the evaluated sample, the dataset size limits broader analysis of the substantial per-image variation. In particular, some images retain watermark traces with BA values above 0.70 after attack. Larger-scale experiments could investigate which image characteristics, objects, or visual features are associated with these residual watermark traces. They could also help determine which image characteristics are associated with discrepancies between LPIPS and qualitative visual quality.

Finally, future watermarking schemes could investigate embedding information across multiple complementary frequency ranges to increase the difficulty of watermark removal while maintaining low perceptual cost. The SSNSW experiments show that combining two watermarks associated with different frequency characteristics can require two attacks for effective removal. Future work should therefore investigate whether broader spectral regions can support additional watermark representations while maintaining visual quality. The residual traces observed after sequential attacks also motivate a more detailed spectral analysis to determine where surviving watermark information remains and whether these components can be exploited to design more robust multi-layer watermarking schemes.

Conclusions

The application of SSNSW schemes increases robustness against single UnMarker attacks. When HiDDeN and StegaStamp are embedded sequentially, regardless of order, both watermarks remain decodable, and the weakening of the initially embedded watermark after embedding the second one is minimal. Embedding order therefore has only a limited effect, with StegaStamp followed by HiDDeN providing slightly better post-embedding decodability for the evaluated pair.

More importantly, neither UnMarker attack can reduce both watermark BA scores close to their respective baselines. The semantic attack offers more disruption to both watermarks, but not enough to remove major traces of the non-semantic scheme. Sequential attacks are therefore required to substantially reduce both watermarks, while attack order has a limited and configuration-dependent effect.

SSNSW therefore raises the effort required for watermark removal but is not a complete solution to adversarial attacks. The study also finds that LPIPS, despite remaining low, can correspond to clear visible distortions, especially when facial images are considered. Future work should be more multi-layered to increase robustness, evaluate additional watermarking and attack schemes, and use better visual metrics alongside larger-scale experiments.

Acknowledgment

The author thanks Dr. Daniel Tenbrinck for his supervision and guidance during the master’s thesis research on which this work is based.

The author also gratefully acknowledges Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) for providing access to the high-performance computing (HPC) resources used to conduct the experiments in this study.

Md Ifraham Iqbal received the B.Sc. degree in Computer Science and Engineering from United International University, Dhaka, Bangladesh, in 2020, and the M.Sc. degree in Data Science from Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany, in 2026. His research has covered machine learning, deep learning, computer vision, health informatics, robotics, and tabular-data learning. His current research interests include defensive image watermarking, watermark-removal robustness, computer vision, and robust AI systems. He has authored and coauthored research published in IEEE Access, IEEE Region 10 Symposium (TENSYMP), and Springer proceedings, with work spanning convolutional neural networks for tabular data, disease prediction, intelligent transportation, and robotics.

References

  1. [1]F. Laurentius and T. Laurentius, Italian Watermarks 1750-1860. Brill, 2016, vol. 50.
  2. [2]Regula Forensics. (2024, Jul.) Banknote watermarks explained: Classification and verification techniques. Accessed: 2026-03-27. [Online]. Available: https://regulaforensics.com/blog/banknote-security-features-watermarks/
  3. [3]J. Karaganis, Media piracy in emerging economies. Lulu.com, 2011.
  4. [4]R. G. Van Schyndel, A. Z. Tirkel, and C. F. Osborne, “A digital watermark,” in Proceedings of 1st international conference on image processing, vol. 2. IEEE, 1994, pp. 86–90.
  5. [5]S. M. Youssef, A. A. ElFarag, and N. M. Ghatwary, “Adaptive video watermarking integrating a fuzzy wavelet-based human visual system perceptual model,” Multimedia Tools and Applications, vol. 73, no. 3, pp. 1545–1573, 2014.DOI
  6. [6]P. Malakar and M. Leeldharan, “Generative ai tools for collaborative content creation: A comparative analysis,” DESIDOC Journal of Library & Information Technology, vol. 44, no. 3, pp. 151–157, 2024.
  7. [7]Y. Huang, “The anatomy of a deepfake voice phishing attack: How ai-generated voices are powering the next wave of scams,” Group-IB Blog, Aug. 2025, accessed 2026-02-06. [Online]. Available: https://www.group-ib.com/blog/voice-deepfake-scams/
  8. [8]Corsearch, “Trade in counterfeit goods market set to reach $1.79 trillion in 2030,” n.d., accessed 2026-02-06. [Online]. Available: https://corsearch.com/about/press-releases/trade-in-counterfeit-goods-market-set-to-reach-1-79-trillion-in-2030/
  9. [9]R. Khan, M. Taqi, and A. Afzal, “Deepfakes in finance: Unraveling the threat landscape and detection challenges,” in Navigating the World of Deepfake Technology. IGI Global, 2024, pp. 91–120.DOI
  10. [10]B. Colman, “Why detecting dangerous ai is key to keeping trust alive in the deepfake era,” World Economic Forum, Jul. 2025, accessed 2026-02-06. [Online]. Available: https://www.weforum.org/stories/2025/07/why-detecting-dangerous-ai-is-key-to-keeping-trust-alive/
  11. [11]E. Eareghan, “From data breaches to deepfakes: a comprehensive review of evolving cyber threats and online risk management,” Communication in Physical Sciences, vol. 9, no. 4, pp. 538–553, 2023.
  12. [12]Financial Post and GlobeNewswire, “Deepfake fraud costs the financial sector an average of $600,000 for each company, regula’s survey shows,” n.d., accessed 2026-02-06. [Online]. Available: https://financialpost.com/globe-newswire/deepfake-fraud-costs-the-financial-sector-an-average-of-600000-for-each-company-regulas-survey-shows
  13. [13]S. Lalchand, V. Srinivas, B. Maggiore, and J. Henderson, “Generative ai is expected to magnify the risk of deepfakes and other fraud in banking,” Deloitte Insights, May 2024, accessed 2026-02-06. [Online]. Available: https://www.deloitte.com/us/en/insights/industry/financial-services/deepfake-banking-fraud-risk-on-the-rise.html
  14. [14]J. Artaius. (2019, Apr.) 2.5 billion online images stolen every day in 2018. [Online]. Available: https://www.digitalcameraworld.com/news/25-billion-online-images-stolen-every-day-in-2018
  15. [15]P. Meliana. (2025, Apr.) When ai fails in public: what brand trust really costs. [Online]. Available: https://www.contentgrip.com/ai-brand-trust-crisis/
  16. [16]Grand View Research. (2025) Digital watermarking market (2025–2033): Size, share & trends analysis report by type (invisible watermarking, visible watermarking), by application, by deployment (on-premises, cloud), by end use, by region and segment forecasts. [Online]. Available: https://www.grandviewresearch.com/industry-analysis/digital-watermarking-market-report
  17. [17]European Parliament. (2025, Feb.) Eu ai act: first regulation on artificial intelligence. [Online]. Available: https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence
  18. [18]The American Presidency Project. (2023, Oct.) Executive order 14110—safe, secure, and trustworthy development and use of artificial intelligence. [Online]. Available: https://www.presidency.ucsb.edu/documents/executive-order-14110-safe-secure-and-trustworthy-development-and-use-artificial
  19. [19]L. Zhang. (2023, Apr.) China: Provisions on deep synthesis technology enter into effect. [Online]. Available: https://www.loc.gov/item/global-legal-monitor/2023-04-25/china-provisions-on-deep-synthesis-technology-enter-into-effect/
  20. [20]W. Zhu, Z. Xiong, and Y-Q. Zhang, “Multiresolution watermarking for images and video,” IEEE transactions on circuits and systems for video technology, vol. 9, no. 4, pp. 545–550, 1999.DOI
  21. [21]J. Zhang and A. T. Ho, “An efficient digital image-in-image watermarking algorithm using the integer discrete cosine transform (intdct),” in Fourth International Conference on Information, Communications and Signal Processing, 2003 and the Fourth Pacific Rim Conference on Multimedia. Proceedings of the 2003 Joint, vol. 2. IEEE, 2003, pp. 1163–1167.DOI
  22. [22]M. S. Raval and P. P. Rege, “Discrete wavelet transform based multiple watermarking scheme,” in TENCON 2003. Conference on Convergent Technologies for Asia-Pacific Region, vol. 3. IEEE, 2003, pp. 935–938.DOI
  23. [23]S. Baluja, “Hiding images in plain sight: Deep steganography,” Advances in neural information processing systems, vol. 30, 2017.
  24. [24]J. Zhu, R. Kaplan, J. Johnson, and L. Fei-Fei, “Hidden: Hiding data with deep networks,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 657–672.arxiv.org/abs/1807.09937
  25. [25]B. Wen and S. Aydore, “Romark: A robust watermarking system using adversarial training,” arXiv preprint arXiv:1910.01221, 2019.arxiv.org/abs/1910.01221
  26. [26]M. Ahmadi, A. Norouzi, N. Karimi, S. Samavi, and A. Emami, “Red-mark: Framework for residual diffusion watermarking based on deep networks,” Expert Systems with Applications, vol. 146, p. 113157, 2020.
  27. [27]K. A. Zhang, L. Xu, A. Cuesta-Infante, and K. Veeramachaneni, “Robust invisible video watermarking with attention,” arXiv preprint arXiv:1909.01285, 2019.arxiv.org/abs/1909.01285
  28. [28]M. Tancik, B. Mildenhall, and R. Ng, “Stegastamp: Invisible hyperlinks in physical photographs,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 2117–2126.arxiv.org/abs/1904.05343
  29. [29]N. Yu, V. Skripniuk, S. Abdelnabi, and M. Fritz, “Artificial fingerprinting for generative models: Rooting deepfake attribution in training data,” in Proceedings of the IEEE/CVF International conference on computer vision, 2021, pp. 14448–14457.
  30. [30]N. Yu, V. Skripniuk, D. Chen, L. S. Davis, and M. Fritz, “Responsible disclosure of generative models using scalable fingerprinting,” in International Conference on Learning Representations, 2022.
  31. [31]S. Gowal, R. Bunel, F. Stimberg, D. Stutz, G. Ortiz-Jimenez, C. Kouridi, M. Vecerik, J. Hayes, S.-A. Rebuffi, P. Bernard et al., “Synthid-image: Image watermarking at internet scale,” arXiv preprint arXiv:2510.09263, 2025.arxiv.org/abs/2510.09263
  32. [32]S. Datathri, A. See, S. Ghaisas, P.-S. Huang, R. McAdam, J. Welbl, V. Bachani, A. Kaskasoli, R. Stanforth, T. Matejovicova et al., “Scalable watermarking for identifying large language model outputs,” Nature, vol. 634, no. 8035, pp. 818–823, 2024.
  33. [33]N. Lukas and F. Kerschbaum, “{PTW}: Pivotal tuning watermarking for {Pre-Trained} image generators,” in 32nd USENIX Security Symposium (USENIX Security 23), 2023, pp. 2241–2258.arxiv.org/abs/2304.07361
  34. [34]P. Fernandez, G. Couairon, H. Jégou, M. Douze, and T. Furon, “The stable signature: Rooting watermarks in latent diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 22 466–22 477.arxiv.org/abs/2303.15435
  35. [35]Y. Wen, J. Kirchenbauer, J. Geiping, and T. Goldstein, “Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust,” arXiv preprint arXiv:2305.20030, 2023.arxiv.org/abs/2305.20030
  36. [36]D. Cheng, X. Li, W.-H. Li, C. Lu, F. Li, H. Zhao, and W.-S. Zheng, “Large-scale visible watermark detection and removal with deep convolutional networks,” in Pattern Recognition and Computer Vision, ser. Lecture Notes in Computer Science. Springer, Cham, 2018, vol. 11258, pp. 27–40.
  37. [37]T. Dekel, M. Rubinstein, C. Liu, and W. T. Freeman, “On the effectiveness of visible watermarks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2146–2154.DOI
  38. [38]Y. Liu, Z. Zhu, and X. Bai, “Wnet: Watermark-decomposition network for visible watermark removal,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2021, pp. 3685–3693.arxiv.org/abs/2012.07616
  39. [39]X. Cun and C.-M. Pun, “Split then refine: stacked attention-guided resunets for blind single image visible watermark removal,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 2, 2021, pp. 1184–1192.
  40. [40]Y. Hu, Z. Jiang, M. Guo, and N. Gong, “Stable signature is unstable: Moving image watermark from diffusion models,” arXiv preprint arXiv:2405.07145, 2024.arxiv.org/abs/2405.07145
  41. [41]Z. Jiang, J. Zhang, and N. Z. Gong, “Evading watermark based detection of ai-generated content,” in Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, 2023, pp. 1168–1181.arxiv.org/abs/2305.03807
  42. [42]M. Saberi, V. S. Sadasivan, K. Rezaei, A. Kumar, A. Chegini, W. Wang, and S. Feizi, “Robustness of ai-image detectors: Fundamental limits and practical attacks,” arXiv preprint arXiv:2310.00076, 2023.arxiv.org/abs/2310.00076
  43. [43]Y. Liu et al., “Image watermarks are removable using controllable regeneration from clean noise,” arXiv preprint arXiv:2410.05470, 2024.arxiv.org/abs/2410.05470
  44. [44]A. Kassis and U. Hengartner, “Unmarker: a universal attack on defensive image watermarking,” in 2025 IEEE Symposium on Security and Privacy (SP). IEEE, 2025, pp. 2602–2620.DOI
  45. [45]T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision – ECCV 2014. Springer, 2014, pp. 740–755. [Online]. Available: https://www.microsoft.com/en-us/research/publication/microsoft-coco-common-objects-in-context/arxiv.org/abs/1405.0312
  46. [46]Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. [Online]. Available: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.htmlarxiv.org/abs/1411.7766
  47. [47]B. An, M. Ding, T. Rabbani, A. Agrawal, Y. Xu, C. Deng, S. Zhu, A. Mohamed, Y. Wen, T. Goldstein et al., “Benchmarking the robustness of image watermarks,” arXiv preprint arXiv:2401.08573, 2024.arxiv.org/abs/2401.08573
  48. [48]S. Ghazanfari, S. Garg, P. Krishnamurthy, F. Khorrami, and A. Araujo, “R-lpips: An adversarially robust perceptual similarity metric,” arXiv preprint arXiv:2307.15157, 2023.arxiv.org/abs/2307.15157

Paper details

vlnc:W_xdpfad5f
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents