This article explores creating a specialized HyperText Transfer Protocol (HTTP) traffic dataset for multi-class web application classification by topic (domain) to identify business logic vulnerabilities. Each record contains the full text of the HTTP request and response, including headers, body, path, and parameters, which distinguishes the dataset from existing analogs limited to either a single category or a unique resource address on the Internet (Uniform Resource Locator) without any content, and is the first publicly available HTTP traffic dataset with multi-domain markup by categories relevant to the tasks of detecting business logic vulnerabilities. A technique for automated web application scanning is proposed, comprising five sequential stages: defining the composition of subject categories, selecting auxiliary sources and a collection method, preparing an initial list of web resources, multi-level web crawling with preservation of the full content of requests and responses and simulating user actions, as well as post-processing and filtering of the collected data. A comparative study of five classification models was conducted on the collected data: three models based on Term Frequency — Inverse Document Frequency (logistic regression, linear support vector machine, and naive Bayes classifier) and two neural network architectures (the lightweight and fast FastText model and a convolutional neural network adapted for the texts based on Kim’s TextCNN architecture). Web application crawling resulted in a dataset of 53,074 records in Parquet format, covering four categories: e-commerce, games, finance, and social networking and messaging. The best result of model comparison was achieved by the TextCNN model. The dataset and technique are published openly. The limitations of the proposed approach to creating a dataset of HTTP traffic for the task of multi-class classification of web applications by topic are the limitation of the coverage of the dataset to four topic domains, the minimal thematically significant content of publicly accessible pages of applications in the category of social networks and instant messengers, and the introduction of translation artifacts into the feature space of about 30 % of the records.
Deep sequential network design requires choosing the number and placement of residual connections, while architecture-search methods are resource-intensive and do not localize structural defects. The objective is to diagnose pre-training inter-layer coupling decay and use it for minimal residual-topology correction. The procedure analyzes off-diagonal Hessian blocks and introduces a global collapse index and a maximum local drop. A greedy algorithm places residual connections based on the profile without training or gradient-based architecture search. The method is evaluated on multilayer perceptrons with an artificial bottleneck on CIFAR-10. Sixteen configurations with 8, 12, 16, or 20 blocks and base width 64 or 128 were evaluated with three initializations each (48 runs), along with a 16-block architecture of base width 256, bottleneck width 16, and bottleneck position 8. The Spearman correlation between the maximum local drop and final accuracy is –0.692; the p-value is less than 0.001. After architectural adjustment, no significant association with accuracy is detected. With a fixed one-edge budget, guided placement yields higher accuracy than random placement; the p-value is 0.019 (exploratory analysis). Soft Jaccard coefficients range from 0.9583 to 0.9722. For adjacent target blocks, the 90 % confidence interval for the paired accuracy difference lies within the equivalence bounds; the larger p-value of the two one-sided tests is 0.014. The Hessian profile localizes a coupling drop and guides residual intervention rather than serving as an independent predictor of training quality. Conclusions obtained are limited to networks with an artificial bottleneck.
In modeling of steady-state metabolic fluxes in metabolic networks, Flux Balance Analysis (FBA) is widely used. However, the standard FBA formulation can admit biochemically infeasible solutions due to thermodynamically infeasible (futile) cycles. Enforcing thermodynamic feasibility with loopless constraints yields an NP-hard problem, which motivates the development of efficient Mixed-Integer Linear Programming (MILP) formulations of loopless FBA. In this work, we propose a new efficient metabolic potentials method for FBA with loopless constraints. This method is based on a novel MILP formulation of loopless constraints. Our analysis of the prior Fast-SNP method, which achieved an order-of-magnitude speedup, allowed us to identify and quantify the contributions of two key optimization ideas: using sparse constraints and reducing the number of integer variables. The proposed metabolic potentials method adopts these ideas and introduces an equivalent MILP formulation of loopless constraints. We compare the performance of the metabolic potentials method using metabolic models from the BiGG database. The proposed method shows performance comparable to Fast-SNP on small and medium scale metabolic models and achieves an approximately 30 % reduction in runtime on a large genome-scale human cell metabolic model RECON1. We also demonstrate that reducing the number of integer variables is the dominant factor in the overall speedup. The proposed metabolic potentials method retains the benefits of the optimization ideas introduced in Fast-SNP and achieves the same or better efficiency. Therefore, it can be used as a drop-in replacement for Fast-SNP and other standard methods for FBA with loopless constraints.
A method for estimating insertion losses introduced by spectrally selective elements is proposed, which enables an increase in the dynamic range of measurements and subsequent calculations of the performance characteristics for a Continuous Variable Quantum Key Distribution system. In contrast to common approaches that employ supercontinuum sources, the proposed method provides more accurate assessment of the system’s resilience against the Trojanhorse attack. The essence of the method consists in using a tunable laser source with additional filtering incorporated into the optical setup for insertionloss measurements. The paper presents a comparative analysis of loss measurement results obtained with a supercontinuum source and a tunable laser source. For the tunable source, the addition of extra filtering in the optical scheme is proposed. It is demonstrated that the maximum loss values measured with the tunable source exceed those obtained with the supercontinuum source by 20 dB. The experiments reveal that employing the tunable source together with additional filtering enhances the dynamic range of insertionloss measurements for Coarse Wavelength Division Multiplexing filters. Using the Trojanhorse attack as a case study, it is shown that additional filtering in the tunablesource configuration yields more accurate transmittance values. This, in turn, affects the formulation of requirements for the implementation of countermeasures, aiming to maintain the quantum key generation rate as close as possible to the case without the attack.
In modern radiology, diagnostic Artificial Intelligence assistants for classifying Magnetic Resonance Imaging (MRI) brain images are typically based on Convolutional Neural Networks (CNNs), which work exclusively with the visual signal and are unable to use structured diagnostic descriptions of pathologies. Vision Language Models (VLMs) open up new possibilities: joint processing of visual and textual data allows for the consideration of the semantic context of pathologies, which is inaccessible to CNN classifiers. The key obstacle to their application remains the high resource intensity of retraining, which is incomparable with the capabilities of a typical clinic. This paper proposes a method for constructing a resource-efficient diagnostic AI assistant based on VLM Gemma3 (4b-it) using low-rank adaptation (LoRA). Retraining was performed on a dataset of 91,777 Fluid-Attenuated Inversion Recovery MRI brain images, distributed across eight pathology classes. Only low-rank adapters were trained — 12.5 % of the model parameters, which allowed training on a single NVIDIA RTX3090 Graphics Processing Unit with 24 GB of Video Random Access Memory without masked data annotation. The radiologist’s expert knowledge was integrated via a system prompt, which assigns the model the role of an experienced neuroradiologist with explicit analysis criteria. On the validation set (9178 examples), the model achieved: Accuracy 0.930, F1-score 0.929, Specificity 0.989, and Cohen’s Kappa 0.915. Comparison with the baseline ConvNeXt-Base classifier on an identical dataset showed the superiority of VLM+LoRA across all metrics: Accuracy increased by 7.4 percentage points, and Cohen’s Kappa by 0.090. The obtained results confirm the practical applicability of the proposed approach for clinics with limited computing resources.
