White Box Watermarking for Convolution Layers in Fine-Tuning Model Using the Constant Weight Code.

Area of Science:

• Artificial Intelligence
• Computer Science
• Cybersecurity

Background:

• Deep neural network (DNN) watermarking protects intellectual property but faces challenges like neuron pruning and limited embedding layers.
• Existing methods often focus on robustness against retraining and fine-tuning, with watermarks typically embedded only in fully connected layers.

Purpose of the Study:

• To extend DNN watermarking techniques to be applicable to any convolution layer within a DNN model.
• To develop a robust watermark detector using statistical analysis of extracted weight parameters.
• To leverage nonfungible tokens for enhanced watermark security and timestamping.

Main Methods:

• Developed an extended DNN watermarking method applicable to convolution layers.
• Designed a watermark detector based on statistical analysis of weight parameters.
• Integrated nonfungible token technology to prevent watermark overwriting and record creation time.

Main Results:

• The proposed method successfully embeds watermarks in convolution layers, expanding applicability beyond fully connected layers.
• The statistical watermark detector effectively identifies the presence of watermarks.
• Nonfungible tokens provide a secure and verifiable method for protecting DNN watermarks.

Conclusions:

• This research significantly advances DNN watermarking by enabling watermark embedding in any convolution layer.
• The integration of statistical analysis and nonfungible tokens offers a more robust and secure solution for protecting DNN models.
• The findings contribute to safeguarding intellectual property rights in the field of artificial intelligence.