Privacy-Preserving Deep Learning in Computational Social Systems via Fully Homomorphic Encryption

Privacy-preserving deep learning is crucial for computational social systems such as healthcare and finance, where sensitive high-dimensional data (e.g., images, video, and sensor streams) demands robust confidentiality and model security. Traditional privacy methods often sacrifice accuracy or scalability through simplifications such as dimensionality reduction and polynomial approximations. To overcome these challenges, this article introduces the multilinear tensor product cryptosystem (MTPC), a novel fully homomorphic encryption (FHE) framework. MTPC enables efficient and accurate encrypted inference on deep neural networks with high-dimensional inputs by constructing structured multilinear mappings over ideal lattices. This approach transforms complex operations such as convolutions and nonlinear activations into encrypted tensor contractions, effectively bypassing the quadratic multiplicative depth limitations of conventional FHE schemes. Experimental evaluations on benchmark medical datasets demonstrate a minimal mean calculation error of only 0.87%. MTPC offers a scalable, trustworthy, and practical foundation for secure artificial intelligence (AI) deployment in privacy-sensitive computational social systems.