Defense Method Against Adversarial Example Attacks using Thermal Noise of a CMOS Image Sensor

With the development of IoT technology, edge AI has attracted widespread attention. However, edge AI shows vulnerability to adversarial example (AE) attacks, where invisible small perturbations can cause a convolutional neural network (CNN) to misclassify input images. Defense methods that add disturbance noise to the input images have been proposed as a solution to AE attacks on edge AI. In this paper, we first show that the defense method using disturbance noise is less effective when the noise generated (e.g., with an arithmetic algorithm) is predictable. Therefore, we propose a defense method using unpredictable noise based on thermal noise generated by an image sensor, along with a post-processing method for the thermal noise. An image sensor was utilized to generate 5-bit disturbance noise for defense against AE attacks based on thermal noise. Using the NIST 800–22 random number test, all items passed with high probability, indicating high randomness. It is also confirmed that the defense performance was improved by 24% when using the proposed post-processing approach.