A Deep Neural Network Guided Testing Approach for Finite State Machines

Finite-State Machines (FSMs) lie at the heart of every digital system and verifying whether a given circuit implementation (B) of an FSM (A) conforms to its specification is an important task in the design cycle. In this work, a deep neural network (DNN) based technique for testing FSMs is developed. Given a set of transition functions that specify an FSM, a DNN is trained with the input-output sequences using the back-propagation algorithm. First, the input sequences and the corresponding output sequences (I/O-pairs) for A are constructed, and some of them are utilized to train the DNN. After training, the proposed DNN is validated with the rest of the I/O-pairs. Once the training and validation of the DNN is completed, it can be used for checking the correctness of its implementation (B) very quickly. Some inputs are applied to B and the observed output sequences are compared with those predicted by the proposed DNN. Based on the similarities between them, B is either declared as correct implementation of A, else it is declared as faulty implementation. The experiments are performed on the MCNC FSM benchmarks and certain faults are injected to form mutant FSMs. Experimental results reveal the efficacy of the proposed technique. Only a few tests are required to detect the presence of anomaly, if any. Hence, the test time is very less resulting in an average test time reduction of 85.66% compared to existing method. It is observed from the earlier works that this type of DNN based FSM testing method is presented first time.