On the Limitations of Logic Locking the Approximate Circuits

Approximate Computing (AxC) is an emerging architectural paradigm, which promises improvements in area, performance or power consumption at the expense of limited and tolerable inaccuracies of computations. Circuits designed in typical AxC-friendly applications, such as neural network accelerators, visual computing and high-bandwidth communication, are vulnerable to supply-chain attacks. In this paper, we investigate logic locking-a protective technique against counterfeiting and overbuilding-in the context of AxC arithmetic primitives. We discuss a threat model that can hinder the adoption of logic locking on AxC based ICs. We show that modern SAT-resilient locking techniques are unsuitable for approximate circuits because of the low output corruption problem. Moreover, we explore the resistance of logic locking against an attacker who managed to obtain a subset of the correct locking key's bits. We show that an incorrectly unlocked circuit can still produce acceptable results, experimentally verifying this both at the level of primitives (adders and multipliers) and in a full application (neural network inference with approximate multiply-accumulate units).