# Backdoors in Conditional Diffusion: Threats to Responsible Synthetic Data Pipelines

arXiv: https://arxiv.org/abs/2507.04726 · HTML: https://mlsanity.com/paper/2507.04726/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Robustness, privacy and adversarial aspects, one of 14 directions in Diffusion models.
- Builds on 19 of its 32 references that are in the map.
- First posted on arXiv 7 Jul 2025.

## Placed in

- Diffusion models / [Robustness, privacy and adversarial aspects](https://mlsanity.com/diffusion/robustness-privacy-and-adversarial-aspects/)
- LLM safety and security / [Backdoor and data poisoning attacks](https://mlsanity.com/safety/backdoor-and-data-poisoning-attacks/)

## Builds on (most cited first, 19 of 19)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Denoising Diffusion Probabilistic Models](https://mlsanity.com/paper/2006.11239/) (2020)
- [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://mlsanity.com/paper/2307.01952/) (2023)
- [Deep Learning Face Attributes in the Wild](https://mlsanity.com/paper/1411.7766/) (2014)
- [Progressive Growing of GANs for Improved Quality, Stability, and Variation](https://mlsanity.com/paper/1710.10196/) (2017)
- [OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields](https://mlsanity.com/paper/1812.08008/) (2018)
- [BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain](https://mlsanity.com/paper/1708.06733/) (2017)
- [Extracting Training Data from Diffusion Models](https://mlsanity.com/paper/2301.13188/) (2023)
- [Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks](https://mlsanity.com/paper/1804.00792/) (2018)
- [Spectral Signatures in Backdoor Attacks](https://mlsanity.com/paper/1811.00636/) (2018)
- [Invisible Backdoor Attack with Sample-Specific Triggers](https://mlsanity.com/paper/2012.03816/) (2020)
- [Weight Poisoning Attacks on Pre-trained Models](https://mlsanity.com/paper/2004.06660/) (2020)
- [Poisoning Web-Scale Training Datasets is Practical](https://mlsanity.com/paper/2302.10149/) (2023)
- [Text-to-Image Diffusion Models can be Easily Backdoored through Multimodal Data Poisoning](https://mlsanity.com/paper/2305.04175/) (2023)
- [Nightshade: Prompt-Specific Poisoning Attacks on Text-to-Image Generative Models](https://mlsanity.com/paper/2310.13828/) (2023)
- [Diffusion Models Beat GANs on Image Classification](https://mlsanity.com/paper/2307.08702/) (2023)
- [Auditing and Generating Synthetic Data with Controllable Trust Trade-offs](https://mlsanity.com/paper/2304.10819/) (2023)
- [Opportunities and Challenges of Frontier Data Governance With Synthetic Data](https://mlsanity.com/paper/2503.17414/) (2025)
- [Deepfake Forensic Analysis: Source Dataset Attribution and Legal Implications of Synthetic Media Manipulation](https://mlsanity.com/paper/2505.11110/) (2025)

## Built on by (most cited first, 1 of 1)

- [Breaking Audio Large Language Models by Attacking Only the Encoder: A Universal Targeted Latent-Space Audio Attack](https://mlsanity.com/paper/2512.23881/) (2025)

## Abstract (arXiv)

Text-to-image diffusion models achieve high-fidelity image generation from natural language prompts. ControlNets extend these models by enabling conditioning on structural inputs (e.g., edge maps, depth, pose), providing fine-grained control over outputs. Yet their reliance on large, publicly scraped datasets and community fine-tuning makes them vulnerable to data poisoning. We introduce a model-poisoning attack that embeds a covert backdoor into a ControlNet, causing it to produce attacker-specified content when exposed to visual triggers, without textual prompts. Experiments show that poisoning only 1% of the fine-tuning corpus yields a 90-98% attack success rate, while 5% further strengthens the backdoor, all while preserving normal generation quality. To mitigate this risk, we propose clean fine-tuning (CFT): freezing the diffusion backbone and fine-tuning only the ControlNet on a sanitized dataset with a reduced learning rate. CFT lowers attack success rates on held-out data. These results expose a critical security weakness in open-source, ControlNet-guided diffusion pipelines and demonstrate that CFT offers a practical defense for responsible synthetic-data pipelines.