An Orthogonal Feature Space as a Watermark: Harmless Model Ownership Verification by Watermarking Feature Weights

High-performance deep learning models require extensive computational resources and datasets, making their ownership protection a pressing concern. To address this challenge, we focus on advancing model security through robust watermarking mechanisms. In this work, we propose a novel deep neural network watermarking method that embeds ownership information directly within the image feature space. Unlike existing approaches that often suffer from low embedding success rates and significant performance degradation, our method leverages convolutional kernels with orthogonal preferences to extract multiperspective features, which are then linearly mapped at the output layer for watermark embedding. Furthermore, we introduce an orthogonal regularization constraint into the loss function to increase the watermark robustness. This constraint enforces orthogonality in both convolutional and fully connected layer weights, suppresses redundancy in hidden layer representations, and minimizes interference between the watermark and the model’s original feature space. Through these innovations, we significantly improve the embedding reliability and preserve model integrity. Experimental results obtained on ResNet-18 and ResNet-101 demonstrate a 100% watermark detection rate with less than 1% performance impact, underscoring the practical security value of our approach. Comparative analysis further validates that our method achieves superior harmlessness and effectiveness relative to state-of-the-art techniques. These contributions highlight the role of our work in strengthening intellectual property protection and the trustworthy deployment of deep learning models.