Singularization: An Efficient Alternative to AES for Safeguarding Model Weights

In the era of Machine Learning, the security of trained models using Deep Learning techniques has become a critical concern, particularly in scenarios where models are deployed in untrusted environments. This paper presents a novel approach to safeguarding the weights of the neural network: singularization, which involves random permutations of the weight matrices. A comparison of the performance and security of singularization with standard AES encryption is made, demonstrating that singularization not only offers a faster alternative but also maintains a level of protection suitable for practical applications. It is shown that singularization can serve as a lightweight protection mechanism for model weights, enabling secure storage without the computational overhead associated with conventional encryption methods.