ME: Trigger Element Combination Backdoor Attack on Copyright Infringement

Aishan Liu(刘艾杉)arXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a new backdoor attack called Multi-Element (ME) combined with Discrete Cosine Transform (DCT) to exploit copyright infringement vulnerabilities in diffusion models like Stable Diffusion. By creating more complex and stealthy poisoned samples, the attack significantly improves the success rate, especially in low-data scenarios where previous methods failed.

What Problem Does This Paper Address?

Generative diffusion models, like Stable Diffusion, are incredibly good at replicating patterns from their training data. This capability can be exploited to create "copyright infringement attacks," where a model is subtly trained to generate images that are similar to specific copyrighted works when prompted with certain triggers. While methods like SilentBadDiffusion (SBD) exist, they face challenges:

  • Limited Data: Creating effective poisoned datasets is difficult due to copyright concerns, inappropriate content, and the unsuitability of many existing images for attack methods.
  • Low Effectiveness in Low-Data Scenarios: Existing methods struggle when only a small fraction of poisoned samples make it into the training data.
  • Stealthiness: Attacks need to be subtle enough not to be detected during training.

This paper aims to address these limitations by proposing new datasets and a more robust, stealthy attack method.

Methods

The paper introduces two key advancements:

  • Multi-Element (ME) Attack: Instead of using a single trigger element (like a specific word or visual pattern), ME combines multiple visual and text elements within a single poisoned sample. This increases the complexity and potential impact of the trigger.
  • Discrete Cosine Transform (DCT): To ensure stealthiness, DCT is applied to the poisoned samples. DCT transforms image data into frequency components, allowing the researchers to selectively retain high-frequency information while discarding lower-frequency details. This process reduces the visual similarity between the poisoned sample and the original target image, making it harder to detect, while still preserving enough information for the attack.

These methods are applied to the SBD framework, enhancing its effectiveness and stealth.

Datasets

The authors introduce two new datasets for evaluating copyright infringement attacks:

  • Style: A synthesized dataset of 60,000 images generated in various artistic styles (aquarelle, frosting lane, half illustration, PS1, tarot, yarn) using prompts from MS COCO2017.
  • DiffusionDB: A dataset of 14 million images generated by Stable Diffusion.

For experiments, subsets of 800 images were used from each new dataset, with one image designated as the target for copyright infringement and 600 images as clean data. They also used the Midjourney dataset for comparison.

Key Insights and Ideas

  • Complexity Enhances Attack: Combining multiple trigger elements (ME) significantly boosts attack effectiveness, especially when the number of poisoned samples is limited.
  • DCT for Stealth: DCT is a crucial technique for making poisoned samples less visually similar to the original target, thus improving stealth without sacrificing attack capability.
  • New Datasets are Crucial: The proposed Style and DiffusionDB datasets are shown to be suitable for evaluating these attacks, with DiffusionDB performing particularly well in terms of attack success rate.
  • Low-Data Scenarios are Vulnerable: Even with a very low subsampling ratio (e.g., 5% poisoned samples), the proposed ME+DCT attack remains highly effective, outperforming previous methods that failed entirely.

Key Results

  • On the new Style and DiffusionDB datasets, the ME+DCT attack achieved Copyright Infringement Rates (CIR) of 16.78% / 39.50 and 51.20% / 23.60 respectively, often outperforming benchmark datasets.
  • In extreme low subsampling ratios (5%, 6 poisoned samples), ME+DCT achieved a CIR of 0.23% / 84.00 and 12.73% / 65.50, while the original SBD failed completely.
  • Table 3 shows that as subsampling ratio decreases, ME and DCT methods consistently outperform SBD, with ME showing better performance at higher ratios and DCT at lower ratios.
  • Visualizations (UMAP) and similarity metrics (SSCD, CLIP, DINO) demonstrate that the poisoned samples are stealthy and do not significantly alter the distribution of the training data.

Limitations and Future Work

  • Dataset Suitability: The paper notes that many existing datasets are not ideal for this type of attack due to issues like low data density, unsuitable images for segmentation, or difficulty in hiding poisoning samples stealthily.
  • Attack Method Limitations: While ME+DCT is effective, the attack's success is still influenced by the intensity of visual transformations, trigger design, and the architecture of the diffusion models.
  • Future Directions: The authors suggest exploring other image transformations beyond DCT for increased stealth and potentially more dynamic trigger elements. They also mention the need for further research into more robust defenses against such attacks.

Summary updated

Abstract

The capability of generative diffusion models (DMs) like Stable Diffusion (SD) in replicating training data could be taken advantage of by attackers to launch the Copyright Infringement Attack, with duplicated poisoned image-text pairs. SilentBadDiffusion (SBD) is a method proposed recently, which shew outstanding performance in attacking SD in text-to-image tasks. However, the feasible data resources in this area are still limited, some of them are even constrained or prohibited due to the issues like copyright ownership or inappropriate contents; And not all of the images in current datasets are suitable for the proposed attacking methods; Besides, the state-of-the-art (SoTA) performance of SBD is far from ideal when few generated poisoning samples could be adopted for attacks. In this paper, we raised new datasets accessible for researching in attacks like SBD, and proposed Multi-Element (ME) attack method based on SBD by increasing the number of poisonous visual-text elements per poisoned sample to enhance the ability of attacking, while importing Discrete Cosine Transform (DCT) for the poisoned samples to maintain the stealthiness. The Copyright Infringement Rate (CIR) / First Attack Epoch (FAE) we got on the two new datasets were 16.78% / 39.50 and 51.20% / 23.60, respectively close to or even outperformed benchmark Pokemon and Mijourney datasets. In condition of low subsampling ratio (5%, 6 poisoned samples), MESI and DCT earned CIR / FAE of 0.23% / 84.00 and 12.73% / 65.50, both better than original SBD, which failed to attack at all.
BibTeX
@misc{liu2025me,
  author = {Aishan Liu(刘艾杉)},
  title = {{ME: Trigger Element Combination Backdoor Attack on Copyright Infringement}},
  year = {2025},
  eprint = {2506.10776},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.10776},
}
Figure