Homomorphic encryption

Homomorphic encryption (HE) enables arithmetic computation directly on ciphertext. Data providers encrypt the input using a private key, and the results of computation remain encrypted. Fully homomorphic encryption (FHE) allows arbitrary operations on ciphertext but is less efficient. By leveraging homomorphic encryption, neural networks can compute on encrypted data to protect data privacy. However, homomorphic encryption requires a polynomial representation, whereas neural networks utilize non-linear layers for activation. Some methods approximate non-linear layers with polynomials. Nonetheless, the capability of neural networks relies on non-polynomial activations, leading to reduced accuracy in encrypted models. Moreover, homomorphic encryption introduces extraordinary latency increases, and to date, there's limited work utilizing homomorphic encryption on large models or evaluating it on large datasets.