HANNOVER MESSE 2019, 01 - 05 April
switch to:
Artificial Intelligence

IBM is working on a copy protection for AI systems

Deep learning specialists are expensive and their work is valuable. Researchers at IBM have therefore developed a digital watermark for AI models.

28 Jul. 2018
IBM is working on a copy protection for AI systems (Photo: Maserati)

With Watson , IBM has an AI wizard in their program that can quite easily compete with Amazon Alexa or Google Assistant. However, instead of integrating its product into consumer electronics, such as Amazon's echo cylinders, the Group provides it to industrial customers more or less free of charge. For example, Watson is used for voice control system in the Maserati GranCabrio .

Doing so however carries the risk of other companies taking over technology and incorporating it into their own products. To prevent this, IBM researchers have developed a kind of copy protection for AI systems based on the digital watermark, such as with photos. This should make it possible to determine via a few API queries whether an AI model is watermarked and thus protected or not. The marking itself can be done in three different ways: either you transfer project-specific data and training data into the model or you can add freely selectable data or you add a noise to the processed data. The researchers nevertheless acknowledge that spying on and thus replication of AI cannot be prevented in this way. IBM initially wants to use the procedure internally and later possibly also offer it to customers.