We intend to explore partially homomorphic encryption (PHE) schemes, such as Paillier, which are more efficient than fully homomorphic encryption for specific operations like addition in AI workloads [100]. Research will optimize PHE for matrix operations critical to neural networks, reducing computational costs within Substrate's weight-based execution model. This approach allows computations on encrypted [[ZKP/Data Marketplace/Tokenized Datasets/Comprehensive Mechanisms of Tokenized Datasets|data]] without decryption, ensuring that sensitive data remains protected throughout the process, even during intermediate steps of AI training verified through Substrate's verification infrastructure. For instance, in a neural network, PHE could enable secure addition of encrypted weights, [[ZKP/Data Marketplace/Security and Privacy/Security and Privacy Foundations|preserving privacy]] while maintaining computational efficiency within [[Substrate]]'s runtime environment. See also: [[ZKP/Research/Masking/Hybrid Privacy Models|Hybrid Privacy Models]]