The field of digital watermarking and model provenance is rapidly evolving, with a focus on developing robust and efficient methods for protecting intellectual property and ensuring accountability in AI systems. Recent research has explored new approaches to watermarking, including dual-space smoothing and model editing, which offer improved robustness and stealthiness. Additionally, there is a growing interest in developing methods for provenance tracking and ownership verification, particularly in the context of federated learning and black-box models. Noteworthy papers in this area include DSSmoothing, which proposes a certified dataset ownership verification method for pre-trained language models, and EditMark, which introduces a watermarking method based on model editing. Other notable works include AWARE, which presents an audio watermarking approach with adversarial resistance to edits, and Blackbox Model Provenance via Palimpsestic Membership Inference, which investigates the problem of proving model provenance in black-box settings.
Advances in Digital Watermarking and Model Provenance
Sources
DSSmoothing: Toward Certified Dataset Ownership Verification for Pre-trained Language Models via Dual-Space Smoothing
Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization
EditMark: Watermarking Large Language Models based on Model Editing
Rotation, Scale, and Translation Resilient Black-box Fingerprinting for Intellectual Property Protection of EaaS Models
Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection
Watermark Robustness and Radioactivity May Be at Odds in Federated Learning
AWARE: Audio Watermarking with Adversarial Resistance to Edits
Blackbox Model Provenance via Palimpsestic Membership Inference
LLMs can hide text in other text of the same length.ipynb
Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models