Security of AI Models using Paillier cryptosystem

This paper demonstrates the practical operation of the Paillier cryptosystem in securing direct retrogression conclusion while conserving data sequestration. A customer- garçon armature is used, where sensitive input data is translated on the customer side using Paillier’s cumulative homomorphic encryption and reused on the garçon without revealing the raw values. The translated data is subordinated to a direct retrogression model (with a predefined weight and bias), using homomorphic operations to cipher translated prognostications directly. After decryption, the works corresponded exactly to the expected direct results, thus confirming the correctness of the calculations in the translated space. A performance test carried out on a dataset of 30 numerical values showed an average of the encryption and decryption times of 0.31 seconds and 0.28 seconds respectively. Nevertheless, the performance is quite effective for small to medium-scale datasets. Such a system yields a highly secure sequestration- conserving output for sensitive operations like healthcare analytics or fiscal modelling. The findings indicate that the model keeps computational consistency even when the data is encrypted, thus enabling precise and reliable predictions that do not compromise data security. Also, the simplicity and availability of the perpetration using Python and the PHE library make it a practical choice for real- world deployments.  While presently limited to direct models due to the cumulative nature of the Paillier scheme, the approach presents a strong case for extending sequestration- conserving ways to broader classes of machine literacy algorithms. This exploration highlights the eventuality of homomorphic encryption in enabling secure and secure AI systems, especially in scripts where confidentiality is consummate.