Simple Electromagnetic Analysis Against Activation Functions of Deep Neural Networks

From cloud computing to edge computing, the deployment of artificial intelligence (AI) has been evolving to fit a wide range of applications. However, the security over edge AI is not sufficient. Edge AI is computed close to the device and user, therefore allowing physical attacks such as side-channel attack (SCA). Reverse engineering the neural network architecture using SCA is an active area of research. In this work, we investigate how to retrieve an activation function in a neural network implemented to an edge device by using side-channel information. To this end, we consider multilayer perceptron as the machine learning architecture of choice. We assume an attacker capable of measuring side channel leakages, in this case electromagnetic (EM) emanations. The results are shown on an Arduino Uno microcontroller to achieve high quality measurements. Our experiments show that the activation functions used in the architecture can be obtained by a side-channel attacker using one or a few EM measurements independent of inputs. We replicate the timing attack in previous research by Batina et al., and analyzed it to explain how the timing behavior acts on different implementations of the activation function operations. We also prove that our attack method has the potential to overcome constant time mitigations.