Arithmetic Circuit Classification Using Convolutional Neural Networks

Abstract

Circuit structural recovering is a technique that derives functional blocks using low-level description. Its value resides chiefly in recovering information of high-level description from a project of integrated circuit when its design is lost, Intellectual Property (IP) synthesized in Field-Programmable Gate Array (FPGA), and in others cases when its structure is desirable. Furthermore, many SAT (Boolean satisfiability problem) solvers take advantage of information they can recover and knowledge of the domain problem to improve their processing (e.g. Algorithm Portfolios and Combinational Equivalence Checking). Being so, recovering circuit structural information is a key ingredient for improvements in formal verification using SAT solvers; a helpful step of structural recovering is the functional block identification. Taking an image generated from a circuit’s CNF description, and considering CNN’s maturity on the image recognition domain, we are able to map a macro functional block with a very high accuracy. The main contributions of this paper are the following: (i) We propose the innovative identification of functional blocks through images (ii) We implemented the system based on CNN using TensorFlow (iii) Our experimental results obtained an accuracy over 80%.