TL;DR
This paper introduces PoisonLoRA, the first systematic study of supply-chain vulnerabilities within the text-to-image (T2I) plugin ecosystem. It demonstrates how malicious actors can inject stealthy, robust backdoors into shared Low-Rank Adaptation (LoRA) plugins that achieve nearly 100% attack success rates. Alarmingly, these poisoned plugins propagate "virally," surviving and spreading even when users switch base models or mathematically merge multiple plugins together.
What Problem Does This Paper Address?
The text-to-image generation landscape has shifted from massive, monolithic models to a decentralized, collaborative "share-and-play" ecosystem. Platforms like Civitai and Liblib host hundreds of thousands of user-contributed LoRA plugins—lightweight fine-tunes that allow anyone to inject specific artistic styles or characters into base models like Stable Diffusion.
However, this ecosystem is built on implicit trust. While platforms scan uploaded preview images and text prompts for harmful content, they treat the underlying model weights as a blind spot.
Existing backdoor attacks fail in this environment because they assume a static deployment. In reality, T2I users constantly manipulate plugins: they adjust scaling factors, swap out the underlying base models, and merge multiple LoRAs together to combine styles. Standard backdoors are fragile and quickly break under these perturbations. This paper addresses a critical question: Can an attacker engineer a malicious LoRA plugin that survives the chaotic, user-driven remixing process and virally infects the downstream ecosystem?
Methods
The authors frame plugin poisoning as a parameter-space adaptation problem, utilizing an optimization strategy designed to find "flat minima" in the loss landscape. This ensures that the injected malicious behavior remains stable even when the weights are slightly altered by users.
The framework is instantiated through two distinct attack methods:
- Concept Hijacking (via Poisonous Distillation): Designed for subtle influence operations (e.g., forcing a generic prompt like "sports" to render a specific brand logo, or injecting phishing links). The attacker uses a high-quality "teacher" LoRA to generate a tiny dataset (fewer than 20 samples), then trains a student LoRA using a composite loss that preserves the original style while forcing the hijacked concept association.
- Task Injection (via Attention Steering): Designed to bypass safety filters to generate prohibited content (e.g., NSFW or violent imagery) only when triggered by a secret key (e.g.,
pwd=123). This is a rapid, data-free approach that surgically modifies the key ($K$) and value ($V$) projection matrices within the U-Net's cross-attention layers, mapping the secret trigger directly to the target concept's embedding space without degrading benign performance.
To ensure the payload survives real-world modifications, the authors apply a single-step adversarial ascent during training, forcing the model to optimize against the worst-case parameter perturbations it might encounter during user remixing.
Datasets
- Evaluation Prompts: 50 random captions selected from the MSCOCO dataset, generated 10 times each with different random seeds to evaluate stealthiness and benign utility.
- Custom LoRA Datasets: Built for 7 popular Civitai LoRAs (Artem, Clyde, Line, Bichu, 3DM, Mix4, KoreanDoll). These datasets were constructed using a mix of synthetic data generation (expanding public prompts using ChatGPT to generate 30 high-fidelity image-prompt pairs) and aggregating public showcase images directly from the platforms.
Key Insights and Ideas
- Viral Propagation via Merging: The most striking finding is that creative collaboration can be weaponized. When a user mathematically merges an infected LoRA with a benign one to mix styles, the malicious payload isn't diluted; it copies itself into the new derivative model.
- Robustness Through Flat Minima: Traditional backdoors reside in "sharp" regions of the loss landscape, meaning any weight adjustment destroys the effect. By explicitly optimizing for flatter regions, PoisonLoRA ensures the backdoor survives heavy user modification.
- The Weight Inspection Blind Spot: Current industry defenses focus entirely on surface-level assets (images and text). Because PoisonLoRA remains completely dormant during normal use (exhibiting a near-zero error trigger rate), it easily evades automated platform scanners.
Key Results
- High Success Rates: PoisonLoRA achieved approximately 100% Attack Success Rates (ASR) across 4 distinct real-world scenarios (Phishing Lures, Brand Placement, Sexy Content, and Bloody Content).
- Extreme Robustness: The backdoors maintained near 100% ASR even when transferred to completely different base models (e.g., from DreamShaper to EpiCRealism) and when subjected to different sampling algorithms.
- Chain-Remix Survival: In propagation tests, the malicious payload survived 5 sequential community merges (remixed 5 times with entirely different benign LoRAs) without losing its potency or altering the quality of benign generations.
- Real-World Evasion: The authors successfully uploaded poisoned plugins to both Civitai and Liblib (kept private and deleted immediately after testing), confirming they bypassed all industrial detection pipelines.
Limitations and Future Work
- Threat Scope: The study focused primarily on concept hijacking and task injection. Future work could explore other malicious behaviors, such as injecting subtle societal biases or creating resource-exhaustion vulnerabilities.
- PEFT Generalization: The evaluation was restricted to LoRA plugins. While LoRA is the dominant customization method, the risks associated with other parameter-efficient fine-tuning (PEFT) techniques remain unexplored.
- Advanced Downstream Tweak Testing: The impact of more aggressive user modifications, such as weight pruning or full-parameter secondary fine-tuning, was not fully evaluated.
Summary updated