Research on Gradient-Based Adversarial Attack Methods for Image Recognition

The advent of the big data era has propelled rapid development in the field of artificial intelligence based on deep neural network models. These models are widely used in scenarios such as facial recognition and autonomous driving. However, experiments have shown that deep neural networks are vulnerable and can be easily attacked by adversarial samples with added perturbations. As a current research hotspot in the field of artificial intelligence, methods of adversarial sample attacks are continuously evolving. The presence of adversarial samples poses safety risks to the application of AI in real-life situations. Therefore, researching the generation methods and principles of adversarial samples to find corresponding defensive measures and enhance the security of deep models is an important task at present. This paper first introduces relevant knowledge about adversarial attacks, focusing on the current trends in image adversarial sample research. It particularly analyzes and lists the current gradient-based adversarial attack methods, and proposes an adaptive gradient algorithm for use in adversarial attacks, which has been proven effective through experimentation. Finally, the paper summarizes the current state of gradient-based algorithms and suggests future research directions.