In this work, we introduce a R esolution- I nvariant, task-agnostic N o-box adversarial A ttack (RINA) on visual neural networks, capable of deceiving any system without prior knowledge of its task, architecture, or input resolution. Unlike previous no-box attacks, which often assume knowledge of the task type or rely on a fixed input resolutions, our approach generalizes across different models by targeting the early layers, making it more robust and applicable in real-world scenarios. We propose a novel adversarial framework that perturbs input images in a way that disrupts fundamental feature extraction while maintaining perceptual similarity for human observers. Our method is evaluated against both convolutional and transformer-based architectures, demonstrating a significant drop in classification accuracy, particularly for transformer models, where it outperforms existing no-box attacks such as Hybrid Image Transformation (HIT). Additionally, we provide an open-source implementation 1 to facilitate further research and practical deployment of no-box adversarial testing.

Abstract A new concept, termed virtual neural networks, is introduced, where the count of trainable parameters is kept constant, and scalability is attained purely through computational resources. This concept is an abstract framework that can be realized using any standard convolutional neural network. It merges siamese neural networks with a deep ensemble technique by generating numerous virtual models that share weights derived from a small set of physical models. The ensemble comprises up to hundreds of trained models simultaneously. All virtual networks take the same input, and their interconnected structure induces an internal distortion that boosts the entire ensemble robustness. The accuracy of the ensemble improves as the number of virtual networks increases, without changing the capacity. Virtual neural networks outperform larger capacity models, typical deep ensembles, and contemporary approaches like SWA and Masksembles. Additionally, the highest performing individual model from the ensemble surpasses other models trained individually, even those with a greater number of parameters.

YOLO is a useful, one-stage tool for object detection and classification. In this paper, we consider the application of grocery product detection. The grocery stores have a significant amount of product classes, so it is beneficial to postpone the classification into a second, specialized neural network with a higher capacity. Extracting bounding boxes for a classification network is not precise enough as the detected area includes redundant information about the background. We propose YOLO-ASC, which, for rectangular-based objects, detects bounding boxes together with object contour using a quadrangular. This approach allows detecting objects more accurately and without the background. For the quadrangular detection functionality, YOLO-ASC shares the feature maps that are already present in the network, and therefore its inference time is almost identical to the original YOLO. YOLO reaches high detection precision by using YOLO apriori knowledge, anchors extracted from data. In this work, we present two experiments where we demonstrate that YOLO-ASC training converges faster due to the symbiosis between the bounding box detection and quadrangular detection. Finally, we propose a tool for generating synthetic datasets with quadrangular labels that is helpful for transfer learning.

Although data preprocessing is a universal technique that can be widely used in neural networks (NNs), most research in this area is focused on designing new NN architectures. This paper, we propose a preprocessing technique that enriches the original image data using local intensity information; this technique is motivated by human perception. To encode this information into an image, we introduce a new image structure named image represented by a fuzzy function. When using this structure, a crisp intensity value of each pixel is replaced by a fuzzy set given by a membership function constructed with the usage of extremal values from the particular neighborhood of that pixel. We describe this structure and its properties and propose a way in which it can be used as an input into existing NNs without any modifications. Based on our benchmark consisting of three well-known datasets and five NN architectures, we show that the proposed preprocessing can, in most cases, decrease classification error compared with a baseline and two other preprocessing methods. To support our claim, we have also selected several publicly available projects and tested the impact of the preprocessing with a positive result.

We propose a new hybrid image compression algorithm which combines the F-transform and the JPEG. At first, we apply the direct F-transform and then, the JPEG compression. Conversly, the JPEG decompression is followed by the inverse F-transform to obtain the decompressed image. This scheme brings three benefits: (i) the direct F-transform filters out high frequencies so that the JPEG can reach a higher compression ratio; (ii) the JPEG color quantization can be omitted in order to achieve greater decompressed image quality; (iii) the JPEG-decompressed image is processed by by the inverse F-transform w.r.t. the adjoint partition almost lossless. The paper justifies the proposed hybrid algorithm by benchmarks which show that the hybrid algorithm achieves significantly higher decompressed image quality than the JPEG.

A cheap personal camera is a perfect device that allows us to test algorithms for video enhancement. In this paper we present a cascade of filters based on a fuzzy representation of images. This representation tries to capture the uncertainty underlying in the intensity of a pixel by means of a fuzzy set. The cascade of filters is compared with the corresponding standard filters: blurring, sharpening and image averaging. Besides, experimentally we show that our approach provides similar results with a significant reduction of the computational time.

We solve the problem of network attack detection and classification. We discuss the way of generation and simulation of an artificial network traffic data. We propose an efficient algorithm for data classification that is based on the F-transform technique. The algorithm successfully passed all tests and moreover, it showed ability to perform classification in an on-line regime.

The paper explores Bilinear Interpolation applied to image enlargement after a fuzzification pre-processing. On the one hand, and from a theoretical point of view, we show some interesting relationships between Bilinear Interpolation and the Fuzzification. On the other hand, from an applied point of view we apply the interpolation obtained to enlargement and show that the obtained results are firstly, faster and secondly, comparable with the standard interpolation procedures.

The contribution describes newly developed technique used to improve image quality by fusion of information from the multiple images into one resulting image containing better information than each of the input ones. The presented approach is based on the F-Transform, integral transform used to detect gradients, similarity and image fusion, and noise reduction.