Combine [[ZKP/Research/Advanced Federated Learning/Differential Privacy in Federated Learning|differential privacy]] with [[ZKP/Research/Masking/Homomorphic Masking|homomorphic encryption]] to enable computations on encrypted, [[ZKP/Introduction/About|noisy data,]] adding an extra privacy layer [102] within [[Substrate]]'s runtime environment. This enhances security for [[Introducing Advanced Federated Learning|federated learning.]] Homomorphic encryption allows computations on encrypted data, and adding noise to the outputs ensures that even if the encrypted data is compromised, the noisy results cannot be used to infer individual contributions, providing a robust defense for collaborative AI training across multiple parties coordinated through Substrate's off-chain workers. See also: [[ZKP/Research/Noise/Scalable Noise Addition|Scalable Noise Addition]]