Machine learning applications gain more and more access to highly sensitive information while simultaneously requiring more and more computation resources. Hence, the need for outsourcing these computational expensive tasks while still ensuring security and confidentiality of the data is imminent. In their seminal work, Tramer and Boneh presented the Slalom protocol for privacy-preserving inference by splitting the computation into a data-independent preprocessing phase and a very efficient online phase. In this work, we present a new method to significantly speed up the preprocessing phase by introducing the Carnival protocol. Carnival leverages the pseudo-randomness of the Subset sum problem to also enable efficient outsourcing during the preprocessing phase. In addition to a security proof we also include an empirical study analyzing the landscape of the uniformity of the output of the Subset sum function for smaller parameters. Our findings show that Carnival is a great candidate for real-world implementations.

Almost pairwise independence (API) is a quantitative property of a class of functions that is desirable in many cryptographic applications. This property is satisfied by Learning with errors (LWE)-mappings and by special Substitution-Permutation Networks (SPN). API block ciphers are known to be resilient to differential and linear cryptanalysis attacks. Recently, security of protocols against neural network-based attacks became a major trend in cryptographic studies. Therefore, it is relevant to study the hardness of learning a target function from an API class of functions by gradient-based methods.

We propose a theoretical analysis based on the study of the variance of the gradient of a general machine learning objective with respect to a random choice of target function from a class. We prove an upper bound and verify that, indeed, such a variance is extremely small for API classes of functions. This implies the resilience of actual LWE-based primitives against deep learning attacks, and to some extent, the security of SPNs. The hardness of learning reveals itself in the form of the barren plateau phenomenon during the training process, or in other words, in a low information content of the gradient about the target function. Yet, we emphasize that our bounds hold for the case of a regular parameterization of a neural network and the gradient may become informative if a class is mildly pairwise independent and a parameterization is non-regular. We demonstrate our theory in experiments on the learnability of LWE mappings.

Decentralized Multi-Client Functional Encryption (DMCFE) extends the basic functional encryption to multiple clients that do not trust each other. They can independently encrypt the multiple plaintext-inputs to be given for evaluation to the function embedded in the functional decryption key, defined by multiple parameter-inputs. And they keep control on these functions as they all have to contribute to the generation of the functional decryption keys. Tags can be used in the ciphertexts and the keys to specify which inputs can be combined together. As any encryption scheme, DMCFE provides privacy of the plaintexts. But the functions associated to the functional decryption keys might be sensitive too (e.g. a model in machine learning). The function-hiding property has thus been introduced to additionally protect the function evaluated during the decryption process.

In this paper, we provide new proof techniques to analyze a new concrete construction of function-hiding DMCFE for inner products, with strong security guarantees: the adversary can adaptively query multiple challenge ciphertexts and multiple challenge keys, with unbounded repetitions of the same tags in the ciphertext-queries and a fixed polynomially-large number of repetitions of the same tags in the key-queries. Previous constructions were proven secure in the selective setting only.