Year: 2025
Type: article
Abstract: Artificial intelligence has become widespread due to the effectiveness of solving many applied tasks, including image and video stream processing. However, with the implementation of artificial intelligence technologies in information systems, the number of vulnerabilities is growing. The main information security threats can be implemented by introducing malicious perturbation into input data. The disadvantage of existing methods, which significantly reduce the scope of its application, is binding to the dataset or model architecture. This paper discusses the issues of expanding the scope of methods for L 0 -optimized perturbation elimination as well as the impact of L 0 -optimized attacks on model response. A method of eliminating malicious perturbations embedded by L 0 -optimized attacks on image processing neural networks via statistical transformations is proposed. Image restoration involves the preliminary detection of the perturbation introduced by the attack. Transformations such as bilinear interpolation and Telea algorithm were considered to restore modified pixels. A statistical transformation is applied to the perturbed pixel, after which the restored values of the color components are recorded in its position. Each perturbed pixel is processed sequentially. The restored image is being sent to the input of the protected model. The approach was tested on the CIFAR-10 and MNIST datasets. The restoration method demonstrates a small deviation between the original and restored pixels. There is also an identical response of the model on the restored and original images. It should be noted that response restoration depends on the quality of the model itself. The proposed method is applicable to eliminate perturbations embedded by One-pixel attack and JSMA, however, it does not eliminate adversarial patches. The method is also applicable for color and grayscale images, regardless of the dataset. In addition, the proposed method is potentially universal to the architecture of a neural network and can be used to restore images in a training sample.
Institution: ITMO University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4413636677