Intellectual property protection of neural network architectures via steganography and IEEE 754 standard

This is the authors' abstract. We don't add key points for this paper.

Purpose Neural networks are used in diverse applications, making them vulnerable to tampering and reinforcing the need for ownership authentication. The proposed method is based on a steganographic technique that embeds binary information into the weights using the IEEE 754 representation to enhance the security of the neural network and ownership authentication. Design/methodology/approach The proposed method is assessed using a variational autoencoder. Moreover, this technique can be extended to other neural networks. Ownership information is embedded within the most stable layers of the neural network, determined via gradient-based analysis, to enhance robustness against common model alterations, including fine-tuning, compression, pruning, overwriting, noise injection and weight quantization. Findings The experimental results confirm minimal impact on model performance and ensure reliable data recovery. The bit error rate evaluates the robustness of the proposed method, which obtained values ranging from 0.0131 to 0.129 for different weight pruning (10–50%). These results were further corroborated by extensive experimental validation. Originality/value The proposed method introduces a steganographic technique that embeds ownership information using the IEEE 754 representation. Unlike existing techniques, this approach embeds information into the weights without modifying the model structure and maintains the model’s performance without structural changes.