The ISC International Journal of Information Security

Abstract Machine learning systems, despite exhibiting high inference accuracy in practical applications, are susceptible to security and reliability concerns both during the training phase and the inference phase. In this paper, we have demonstrated that it is possible to extract internal information from a neural network without physical access. This attack was executed through the utilization of a power sensor. This sensor enables remote sampling. Thus far, the sensor has been employed to extract power samples from cryptographic circuits, and its functionality and correctness have been thoroughly tested. Now, in this paper, the same power sensor is used to extract power samples from a neural network, allowing us to assess the supervisor’s performance for applications beyond cryptographic algorithms. In this paper, we demonstrate that the power sensor accurately extracts power samples from neural networks. This paper reveals that between 20,000 and 50,000 power samples of a 16-bit neural network weight can be retrieved. The final step involved hardening the neural network against side-channel attacks. Test results in this section demonstrate that it is possible to make the neural network resistant to first-order side-channel attacks with an area overhead of about 6%. The degree of reinforcement was measured using the assumption test method, revealing that the attack has become eight times more challenging.