Logic obfuscation emerged as an efficient solution to strengthen the security of integrated circuits (ICs) from multiple threats including reverse engineering and intellectual property (IP) theft. Emergence of Boolean Satisfiability (SAT) attacks and its variants have shown to circumvent the security mechanisms such as obfuscation and a plethora of its variants. A plethora of advanced security defenses to thwart the SAT attacks are introduced. Despite the effectiveness, the imposed overheads in terms of area and power are unacceptably high. In contrast, our current work focuses on devising an iterative, dynamic and intelligent SAT-hard clause generator for a given SAT-prone problem, termed as SATConda. The SATConda is a SAT-hard clause generator that utilizes a bipartite propagation based neural network model. The utilized model comprises multiple layers of artificial neural networks to extract the dependencies of literals and variables, followed by long short term memory (LSTM) networks to validate the SAT hardness. The SATConda is trained with conjunctive normal form (CNF) of the IC netlist that are both SAT solvable and SAT-hard. Further, the SATConda is equipped with a SAT-clause generator to convert a CNF from satisfiable (SAT) to unsatisfiable (unSAT) with minor perturbation (which translates to minor overheads) so that the SAT-attack cannot decrypt the keys. To the best of our knowledge, no previous work has
been reported on neural network based SAT-hard clause or CNF translator for circuit obfuscation. We evaluate our proposed SATConda's empirical performance against MiniSAT, Lingeling and Glucose SAT solvers on ISCAS'85 benchmark circuits.