END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions

Authors: Nan Sun, Han Fang, Yuxing Lu, Chengxin Zhao, Hefei Ling

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our scheme outperforms state-of-the-art algorithms under various non-differentiable distortions. Our primary interest is to explore the performance of END2 under fully non-differentiable distortion.
Researcher Affiliation Academia Nan Sun1, Han Fang2, Yuxing Lu3, Chengxin Zhao1, Hefei Ling1* 1 Huazhong University of Science and Technology 2 National University of Singapore 3 Peking University EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1: END2 Pipeline Require: Image-Message pairs (x, m) from dataset, Encoder fθ, Teacher Decoder gt and Parameters δt, Student Decoder gs and Parameters δs, Distortion Function ψ, Momentum Coefficient τ, Optimizer op, Swapping interval k Ensure: Optimized Encoder and Decoders 1: for i, (x, m) in enumerate(dataset) do 2: ˆx fθ(x, m) # Encode image with watermark 3: x sg[ψ(ˆx)] # Apply distortion attack 4: 5: # extract message and feature from encoded images 6: ˆmt, zt gt(ˆx) # Teacher decoder 7: ˆms, zs gs( x) # Student decoder 8: 9: L Ls,t + Lmsg + Lquality # Calculate total loss 10: L.backward() # Backpropagate 11: op.step() # Update model parameters 12: 13: # Momentum updating strategy 14: δt τδt + (1 τ)δs 15: # Swapping Learning strategy 16: if (i + 1) mod k = 0 then 17: gt, gs gs, gt 18: end if 19: end for
Open Source Code No The paper does not provide an explicit statement about the release of source code for the described methodology, nor does it include a link to a code repository.
Open Datasets Yes The model is trained on DIV2K(Agustsson and Timofte 2017). Specifically, we randomly select a block of size 128 128 from the training set as the cover image, and randomly sample a bit stream of length 30 to use as watermarking information for embedding. To evaluate the model s generalization, we also randomly select 2000 images from the COCO (Lin et al. 2014) and Image Net (Deng et al. 2009) datasets for testing.
Dataset Splits Yes Specifically, we randomly select a block of size 128 128 from the training set as the cover image, and randomly sample a bit stream of length 30 to use as watermarking information for embedding. To evaluate the model s generalization, we also randomly select 2000 images from the COCO (Lin et al. 2014) and Image Net (Deng et al. 2009) datasets for testing.
Hardware Specification Yes The batch size is set to 32, and the model is trained for a total of 5000 epochs on a NVIDIA RTX 3090 GPU.
Software Dependencies No The paper mentions using 'PIL package in Python' and 'Adam optimizer' but does not specify exact version numbers for these or other key software components, which is necessary for reproducibility.
Experiment Setup Yes We utilize the Adam optimizer with a fixed learning rate of 8e 4 for training. The batch size is set to 32, and the model is trained for a total of 5000 epochs on a NVIDIA RTX 3090 GPU. ... where λ1, λ2, λ3 are three hyperparameters used to balance different losses, defaulting to 0.01, 8, 5. ... where τ is the weight factor and η is the learning rate, defaulting to 0.999 and 8e 4. ... swapping the two decoders after each k training batches, with k defaulting to 1.