NXP Offers Watermarks for AI Models

Trained machine-learning models can be valuable, and techniques have been developed to identify them in case of theft. NXP has introduced a tool that implements a straightforward approach to watermarking image-classification neural networks.

Anand Joshi

A production-ready neural network requires millions of dollars for training, testing, and deployment, making it an attractive target for thieves. Hackers could profit by reselling it or cause harm by introducing bugs. To thwart such intellectual-property (IP) theft, NXP has introduced a tool that adds a watermark to neural networks.

IP protection can invoke legal means such as patents, trademarks, and copyright. In case of theft, the burden of proof falls on the owner to establish ownership. A watermarking tool can help.

The watermarking tool from NXP is available to customers developing models as a part of its eIQ software development toolset. NXP customers can train and deploy watermarked models on any i.MX application processor.

Although every technology that enhances security is welcome, watermarking won’t stop theft. Neural-network watermarking has undergone academic research but will require more-comprehensive software support to cover the complete model life cycle.