Limits to non-malleability

There have been many successes in constructing explicit non-malleable codes for various classes of tampering functions in recent years, and strong existential results are also known. In this work we ask the following question: When can we rule out the existence of a non-malleable code for a tampering class F? First, we start with some classes where positive results are well-known, and show that when these classes are extended in a natural way, non-malleable codes are no longer possible. Specifically, we show that no non-malleable codes exist for any of the following tampering classes: Functions that change d/2 symbols, where d is the distance of the code; Functions where each input symbol affects only a single output symbol; Functions where each of the n output bits is a function of n − log n input bits. Furthermore, we rule out constructions of non-malleable codes for certain classes F via reductions to the assumption that a distributional problem is hard for F, that make black-box use of the tampering functions in the proof. In particular, this yields concrete obstacles for the construction of efficient codes for NC, even assuming average-case variants of P 6⊆ NC.