Core Concepts
The author proposes AMUSE, an adaptive multi-segment encoding-decoding method for dataset watermarking, aiming to improve message extraction accuracy and watermarked dataset quality.
Abstract
The paper introduces AMUSE, a method that adaptively maps the original watermark into shorter sub-messages to enhance ownership protection of datasets. By distributing the message over multiple samples, AMUSE improves extraction accuracy and dataset quality. The proposed method is evaluated against existing image watermarking techniques through extensive experiments.
Key points:
- Curating high-quality datasets requires effective ownership protection.
- Existing watermarking methods lead to redundancy and reduced quality.
- AMUSE encodes the original message into shorter sub-messages adaptively.
- The encoder adjusts the protection level based on user needs.
- Extensive experiments show improved extraction accuracy and dataset quality with AMUSE.
Stats
Applying AMUSE improves overall message extraction accuracy up to 28%.
Image dataset quality enhanced by approximately 2 dB on average.
Quotes
"The proposed encoder and decoder are plug-and-play modules that can easily be added to any watermarking method."
"AMUSE improves both the extraction accuracy and the quality of the watermarked dataset."