Research map / Adversarial robustness and security / Attacks on text and graphs
MOLM: Mixture of LoRA Markers
Samar Fares, Nurbek Tastan, Noor Hazim Hussein, Karthik Nandakumar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| route LoRA adapters by key in the VAE decoder (MOLM) vs prior bit-recovery watermarks | Stable Diffusion v1.5 · MS-COCO rotation | bit accuracy: 0.56 (Stable Signature) -> 0.84 | Improved |
| MOLM key-routed LoRA instead of Stable Signature fine-tuning | Stable Diffusion v1.5 · MS-COCO | FID: 29.5 -> 27.7 | Improved |
| MOLM vs Stable Signature on JPEG and crop distortions | Stable Diffusion v1.5 · MS-COCO JPEG | bit accuracy: 0.89 vs 0.89 JPEG; crop 0.91 vs 0.97 | No clear effect |
| train the extractor with image augmentations (ablation) | MOLM extractor · Adversarial PGD eps=0.1 | bit accuracy: 0.60 -> 0.96 | Improved |
| train the extractor with augmentation (ablation) | MOLM extractor · Diffusion regeneration, 100 steps | bit accuracy: 0.62 -> 0.82 | Improved |
| perturb the U-Net (108 bits) instead of the VAE decoder (28 bits) (ablation) | MOLM · MS-COCO | image quality (FID): FID increased by 4.15 points | Hurt |
| double capacity with independent bits per convolution (ablation) | MOLM · MS-COCO | bit accuracy: 0.98 -> 0.89 (key 28 -> 56) | Hurt |
| increase LoRA rank from 8 to 64 (ablation) | MOLM · MS-COCO | bit accuracy: 0.75 -> 0.98 | Improved |
| vary sampler, number of steps and CFG scale (ablation) | MOLM · MS-COCO sampling configurations | bit accuracy: 0.94-0.96 | No clear effect |
| average k watermarked images for forgery or removal | MOLM · averaging attacks | bit accuracy: forgery ~0.5; removal >= 0.96 up to 5000 images | No clear effect |
| retrain generator independently as full-knowledge attacker | MOLM extractor · full knowledge attack | bit accuracy: approx 0.5 | No clear effect |
| add 8 paths per layer (3 bits per block) (ablation) | MOLM · MS-COCO | bit accuracy: key 42, bit accuracy 0.90 | Hurt |
Closest papers by meaning
- AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA (ICML 2024 · similarity 0.87)
- BitMark: Watermarking Bitwise Autoregressive Image Generative Models (NeurIPS 2025 · similarity 0.84)
- An Undetectable Watermark for Generative Image Models (ICLR 2025 · similarity 0.84)
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking (NeurIPS 2026 · similarity 0.84)
- CoSDA: Enhancing the Robustness of Inversion-based Generative Image Watermarking Framework (AAAI 2025 · similarity 0.84)
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
