An Approach to Countering Adversarial Attacks on Image Recognition Based on JPEG-Compression and Neural-Cleanse

Nowadays, machine learning is becoming a ubiquitous artificial intelligence technology. It is actively being implemented in various fields of science and technology, including protection against cyber attacks, image recognition, computer vision, autonomous vehicles and other difficult problems. However, despite the advantages of machine learning, this technology is becoming the object of close attention of attackers. Cyber attacks against machine learning models, which can be presented in the form of classifiers or artificial neural networks, can significantly distort the results of these models and cause irreparable damage to a decision-making system based on machine learning. The paper proposes a new approach to countering attacks against machine learning systems in the field of image recognition, which is based on the JPEG-compression and Neural-Cleanse methods. The study shows that the proposed approach is able to restore the ability of a machine learning system to recognize images under gradient adversarial attacks. To demonstrate the effectiveness of the proposed approach, experiments were conducted on four classifiers, which confirmed a significant increase in the accuracy of image recognition after a cyber attack. The results of the work indicate the high potential of using this approach in image recognition systems aimed at detecting and combating adversarial attacks.