AI Paper Atlas

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

ICLR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Neural network watermarking

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 changedWhereMeasuredEffect
route LoRA adapters by key in the VAE decoder (MOLM) vs prior bit-recovery watermarksStable Diffusion v1.5 · MS-COCO rotationbit accuracy: 0.56 (Stable Signature) -> 0.84Improved
MOLM key-routed LoRA instead of Stable Signature fine-tuningStable Diffusion v1.5 · MS-COCOFID: 29.5 -> 27.7Improved
MOLM vs Stable Signature on JPEG and crop distortionsStable Diffusion v1.5 · MS-COCO JPEGbit accuracy: 0.89 vs 0.89 JPEG; crop 0.91 vs 0.97No clear effect
train the extractor with image augmentations (ablation)MOLM extractor · Adversarial PGD eps=0.1bit accuracy: 0.60 -> 0.96Improved
train the extractor with augmentation (ablation)MOLM extractor · Diffusion regeneration, 100 stepsbit accuracy: 0.62 -> 0.82Improved
perturb the U-Net (108 bits) instead of the VAE decoder (28 bits) (ablation)MOLM · MS-COCOimage quality (FID): FID increased by 4.15 pointsHurt
double capacity with independent bits per convolution (ablation)MOLM · MS-COCObit accuracy: 0.98 -> 0.89 (key 28 -> 56)Hurt
increase LoRA rank from 8 to 64 (ablation)MOLM · MS-COCObit accuracy: 0.75 -> 0.98Improved
vary sampler, number of steps and CFG scale (ablation)MOLM · MS-COCO sampling configurationsbit accuracy: 0.94-0.96No clear effect
average k watermarked images for forgery or removalMOLM · averaging attacksbit accuracy: forgery ~0.5; removal >= 0.96 up to 5000 imagesNo clear effect
retrain generator independently as full-knowledge attackerMOLM extractor · full knowledge attackbit accuracy: approx 0.5No clear effect
add 8 paths per layer (3 bits per block) (ablation)MOLM · MS-COCObit accuracy: key 42, bit accuracy 0.90Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy