Deep no-reference tone mapped image quality assessment

Abstract

In ACM CCS�17, Choudhuri et al. designed two fair public-ledger-based multi-party protocols (in the malicious model with dishonest majority) for computing an arbitrary function f. One of their protocols is based on a trusted hardware enclave G

(which can be implemented using Intel SGX-hardware) and a public ledger (which can be implemented using a blockchain platform, such as Ethereum). Subsequently, in NDSS�19, a stateless version of the protocol was published. This is the first time, (a certain definition of) fairness � that guarantees either all parties learn the final output or nobody does � is achieved without any monetary or computational penalties. However, these protocols are fair, if the underlying core MPC component guarantees both privacy and correctness. While privacy is easy to achieve (using a secret sharing scheme), correctness requires expensive operations (such as ZK proofs and commitment schemes). We improve on this work in three different directions: attack, design and performance.

Our first major contribution is building practical attacks that demonstrate: if correctness is not satisfied then the fairness property of the aforementioned protocols collapse. Next, we design two new protocols � stateful and stateless � based on public ledger and trusted hardware that are: resistant against the aforementioned attacks, and made several orders of magnitude more efficient (related to both time and memory) than the existing ones by eliminating ZK proofs and commitment schemes in the design.

Last but not the least, we implemented the core MPC part of our protocols using the SPDZ-2 framework to demonstrate the feasibility of its practical implementation.