The increasing complexity of artificial intelligence (AI) models poses a significant challenge for individuals and organizations without sufficient computing resources to train them. While cloud-based training services can offer a solution, they require sharing sensitive data with untrusted parties, posing risks to data privacy. To address this challenge, we explore the combination of distributed training and homomorphic encryption to parallelize the training process on encrypted data. We utilize the CKKS homomorphic encryption scheme to develop a framework that can train comparably accurate AI models in less time than other homomorphically encrypted training solutions. Our experiments demonstrate reduced total runtime for homomor-phically encrypted model training while maintaining competitive classification accuracy for the MNIST handwritten digits dataset, a well-known benchmarking dataset for machine learning. Our framework brings homomorphic encryption closer to becoming a practical data privacy solution for small stakeholders who cannot afford to compromise on security.

Abstract We describe OCB3, the final version of OCB, a blockcipher mode for authenticated encryption (AE). We prove the construction secure, up to the birthday bound, assuming its underlying blockcipher is secure as a strong-PRP. We study the scheme’s software performance, comparing its speed, on multiple platforms, to a variety of other AE schemes. We reflect on the history and development of the mode.

We describe a parallelizable block-cipher mode of operation that simultaneously provides privacy and authenticity. OCB encrypts-and-authenticates a nonempty string M ε {0,1}• using \lceil |M|/n\rceil + 2 block-cipher invocations, where n is the block length of the underlying block cipher. Additional overhead is small. OCB refines a scheme, IAPM, suggested by Charanjit Jutla. Desirable properties of OCB include: the ability to encrypt a bit string of arbitrary length into a ciphertext of minimal length; cheap offset calculations; cheap session setup; a single underlying cryptographic key; no extended-precision addition; a nearly optimal number of block-cipher calls; and no requirement for a random IV. We prove OCB secure, quantifying the adversary's ability to violate the mode's privacy or authenticity in terms of the quality of its block cipher as a pseudorandom permutation (PRP) or as a strong PRP, respectively.