Lossless Reversible Watermarking for AIGC with Applications in Authentication
Authors: Yupeng Qiu, Han Fang, Ee-Chien Chang
Organizations: School of Computing, National University of Singapore, Singapore · School of Cyber Science and Technology, University of Science and Technology of China, Hefei, Anhui, China
AIGC watermarking is commonly used for tracing, but it can also support applications such as authentication and tampering detection. We study lossless reversible watermarking, which enables exact recovery of the original watermarked image after permissible distortions. By exploiting the relatively low entropy of the information required for image generation, together with robust latent-space watermarking, we design a framework that preserves sufficient information for lossless recovery. We then apply the framework to AIGC authentication, which generally faces two key challenges. First, the authentication mechanism must resist adaptive adversaries. Second, the content relevant to authentication, such as visual designs or expressed messages, may depend on application requirements that are not known at generation time. To solve these difficulties, we incorporate cryptographic message authentication to provide unforgeability against adaptive attacks. Lossless recovery provides the original image as reference evidence, after which a suitable analyzer can be selected to make the authentication decision. Experiments on Stable Diffusion demonstrate exact recovery under a wide range of distortions.
Figures & tables
Figure 1: Lossless Reversible Watermarking (LRW). The sender embeds the compressed prompt, seed, decoding metadata, and HMAC tag as a watermark message. The receiver verifies the extracted message, reconstructs the prompt and watermarked initial noise, and recovers the original image.
Figure 2: LRW-based authentication (LRW-Auth). LRW extracts and verifies the watermark message to recover the original image x . Failed recovery leads to rejection. Otherwise, Compare(y,x) evaluates the received image y against x according to the application’s authentication requirements, returning 1 for acceptance and 0 for rejection.
Figure 3: Distribution of losslessly compressed prompt lengths. We evaluate 67,101 unique SDP prompts and 593,968 MS-COCO prompts. The 1,500-bit threshold accommodates 100% of MS-COCO prompts and 99.83% of SDP prompts.
Figure 4: Visual effects of the evaluated distortions. An LRW (T2SMark) image from SDP and its six transformed versions at the settings in Tables 2 and 3 . Top: full images with the same region marked in red. Middle: fourfold enlargement of that region. Bottom: mean absolute RGB differences from the original region, shown on a shared linear color scale (8-bit intensity units, saturated at 64).
Method
FID ↓
CLIP score ↑
Unwatermarked
20.00
0.3452
LRW (PRC [ 16 ] )
20.49
0.3454
LRW (GS [ 38 ] )
20.43
0.3464
LRW (T2SMark [ 37 ] )
20.51
0.3467
SemBind [ 44 ]
20.51
0.3461
SEAL [ 1 ]
20.66
0.3461
Table 1: Image quality of LRW instantiations. Each LRW row applies the framework using the indicated watermarking method. All methods use the same generation settings.
MS-COCO
SDP
Instantiation
Clean
JPEG compression
Gaussian noise
Gaussian blur
Median filtering
Clean
JPEG compression
Gaussian noise
Gaussian blur
Median filtering
LRW (PRC [ 16 ] )
100%
96%
99%
99%
93%
100%
96%
97%
99%
93%
LRW (GS [ 38 ] )
100%
99%
99%
99%
95%
100%
93%
99%
74%
58%
LRW (T2SMark [ 37 ] )
100%
100%
100%
100%
100%
100%
98%
99%
100%
93%
Table 2: LRW under transmission and processing distortions. Message recovery success rate (%). Success requires complete message recovery and valid MAC verification.
MS-COCO
SDP
Instantiation
Ctrl
Regen
Ctrl
Regen
LRW (PRC [ 16 ] )
95%
99%
95%
97%
LRW (GS [ 38 ] )
96%
98%
63%
74%
LRW (T2SMark [ 37 ] )
100%
100%
98%
99%
Table 3: LRW under diffusion-based purification. Message recovery success rate (%), using the same criterion as Table 2 . Ctrl: CtrlRegen+; Regen: Regen-Diffusion.
Instantiation
MS-COCO
SDP
LRW (PRC [ 16 ] )
97%
96%
LRW (GS [ 38 ] )
98%
89%
LRW (T2SMark [ 37 ] )
99%
99%
Table 4: LRW recovery under reprompting. Complete source-message recovery with valid MAC verification (%).
Method
MS-COCO
SDP
SemBind [ 44 ]
0.9962
0.9698
SEAL [ 1 ]
0.9772
0.9042
LRW-Auth
Forgery resistance by design
Table 5: Forgery detection under reprompting. ROC-AUC ( ↑ ) for SemBind and SEAL. LRW-Auth’s forgery resistance follows from the two-case analysis below.
Figure 5: Reference-based comparison under image editing. A joint identity and logo edit (top) and an object replacement (bottom) illustrate comparison against original references. Brighter regions indicate larger perceptual differences.
Instantiation
InstructPix2Pix
MagicBrush
GPT Image
LRW (PRC [ 16 ] )
86%
72%
64%
LRW (GS [ 38 ] )
86%
83%
70%
LRW (T2SMark [ 37 ] )
90%
91%
76%
Table 6: LRW recovery under instruction-guided editing. Complete source-message recovery with valid MAC verification (%).
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Empirical and fitted compressed-prompt distributions. Histograms of MS-COCO and SDP are shown with normal and lognormal fits, respectively. Bin widths are 5 and 20 bits. Fitting and normalization use all prompts; the SDP display ends at 1,600 bits. The dashed line marks 1,500 bits.
Setting
Value
Generation model
Stable Diffusion 2.1 Base
Image / latent dimensions
512×512 / 4×64×64
Generation solver
DPM-Solver, order 1
Generation steps / guidance
50 / 3
Inversion steps / guidance
50 / 3
Inversion prompt / order
Empty / 0
Appendix
Table S1 : Generation and extraction configuration.
Figure S2 : Empirical ROC curves under reprompting. Original watermarked images are positives and reprompted images are negatives. Thus, the false positive rate is the fraction of reprompted images accepted at a score threshold. Curves use the saved per-image detection scores and reproduce the AUCs in Table 5 .
Figure S3 : Forgery detection by SemBind and SEAL. Rows (a)–(b) compare the two methods on MS-COCO, and rows (c)–(d) on SDP. Each method is shown with one successfully rejected forgery (left) and one accepted forgery (right), alongside their original watermarked images. Detection thresholds are specified in Sec. G.3 .
Statistic
MS-COCO
SDP
Unique prompts
593,968
67,101
Mean
73.09
379.77
SD
16.78
163.42
Median
71
361
95th percentile
102
628
99th percentile
125
840
Appendix
Table S2 : Empirical compressed-prompt statistics. All lengths are in bits; SD is the sample standard deviation.
Figure S4 : Editing a shared original with different editors. Both rows use LRW (T2SMark). Within each row, all three editors receive the same original image and editing instruction. PASS and FAIL denote complete original-message recovery with a valid MAC. Every editor has one successful and one failed recovery.
Diffusion models generate high-quality images but pose serious risks like copyright violation and disinformation. Watermarking is a key defense for tracing and authenticating AI-generated content. However, existing methods rely on threshold-based detection, which only supports fuzzy matching and cannot recover structured watermark data bit-exactly, making them unsuitable for offline verification or applications requiring lossless metadata (e.g., licensing instructions). To address this problem, in this paper, we propose Gaussian Shannon, a watermarking framework that treats the diffusion process as a noisy communication channel and enables both robust tracing and exact bit recovery. Our method embeds watermarks in the initial Gaussian noise without fine-tuning or quality loss. We identify two types of channel interference, namely local bit flips and global stochastic distortions, and design a cascaded defense combining error-correcting codes and majority voting. This ensures reliable end-to-end transmission of semantic payloads. Experiments across three Stable Diffusion variants and seven perturbation types show that Gaussian Shannon achieves state-of-the-art bit-level accuracy while maintaining a high true positive rate, enabling trustworthy rights attribution in real-world deployment. The source code have been made available at: https://github.com/Rambo-Yi/Gaussian-Shannon
Yi Zhang, Hongbo Huang, Liang-Jie Zhang
Center for AI Services Computing, College of Computer Science and Software Engineering, Shenzhen University
Latent-based diffusion model watermarking embeds watermarks into generated images' latent space to enable content attribution, offering a training-free solution for intellectual property protection and digital forensics. However, these methods exhibit a critical vulnerability to the forgery attack, attackers can extract the watermark by inverting the watermarked image and re-generating it with an arbitrary prompt, thereby enabling false attribution on malicious content. In this paper, we propose the CSGuard, the first forgery-resistant watermarking schema that leverages compressed sensing to bind the watermarked image generation and verification to a secret matrix. This ensures that only users possessing the secret matrix can correctly embed or verify the image watermark, prevents the illegal users from forgery without compromising generation quality and watermark integrity. Experimental results demonstrate that CSGuard achieves strong forgery resistance, reduces the attack success rate from 100.0% to 28.12%, and achieve 100% detection rate on benign watermarked images without compromising watermarking effectiveness.
Jiewei Lai, Lan Zhang, Chen Tang +4
University of Science and Technology of China · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center · Lenovo Research
Robust watermarking is critical for intellectual property protection, whereas existing methods face a severe vulnerability against regeneration-based AIGC attacks. We identify that existing methods fail because they entangle the watermark with high-frequency cover texture, which is susceptible to being rewritten during generative purification. To address this, we propose WaterVIB, a theoretically grounded framework that reformulates the encoder as an information sieve via the Variational Information Bottleneck. Instead of overfitting to fragile cover details, our approach forces the model to learn a Minimal Sufficient Statistic of the message. This effectively filters out redundant cover nuances prone to generative shifts, retaining only the essential signal invariant to regeneration. We theoretically prove that optimizing this bottleneck is a necessary condition for robustness against distribution-shifting attacks. Extensive experiments demonstrate that WaterVIB significantly outperforms state-of-the-art methods, achieving superior zero-shot resilience against unknown diffusion-based editing.