Volume 3(68)

CONTENTS

  1. Aleeva V. N., Sapozhnikov A. S. Effcient Implementation of Neural Network Learning Algorithms Using the Concept of a Q-determinant. 
  2. Rahmani J., Baibara B. V., Tetov S. G. Vulnerabilities of Large Language Models: Analysis and Protection Methods.
  3. Malyshkin V. E., Perepelkin V. A., Nushtaev Yu.Yu. Reduction of Invocation Overhead in Automatically Generated Programs with the Active Knowledge Concept. 
  4. Bobokhonov A., Xuramov L., Rashidov A. Detection of Skin Diseases from Images Using Machine Learning and Deep Learning Techniques.
  5. Yurtin A.A. A Method for Forecasting the Error and Training Time of Neural Networks for Multivariate Time Series Imputation.

V. N. Aleeva, A. S. Sapozhnikov

South Ural State University (National Research University), 454080, Chelyabinsk, Russia

EFFICIENT IMPLEMENTATION OF NEURAL NETWORK LEARNING ALGORITHMS USING THE CONCEPT OF A Q-DETERMINANT

DOI: 10.24412/2073-0667-2025-3-5-16

EDN: NGOUCS

In this paper we describe a method for designing Q-effective programs that use the parallelism resource of algorithms completely. This method is used for effective implementation of algorithms.
It has three steps: construction of the Q-determinant of the algorithm, description of the Qeffective implementation of the algorithm, development of a program for an realizable Q-effective implementation of the algorithm. A program is called Q-effective if it is developed using this method. A program is also called Q-effective if it performs a Q-effective implementation of an algorithm. The same set of programs corresponds to these two definitions.

The application of the method of designing Q-effective programs is shown on the example of algorithms implementing stochastic gradient descent and error back propagation methods. These methods are often used to learn neural networks. Q-effective programs for shared and distributed memory parallel computing systems have been developed that implement these methods. The acceleration and efficiency of the developed programs have been evaluated using computational experiments. Computational experiments were performed on the supercomputer «Tornado» of the South Ural State University. We present conclusions based on the obtained evaluation of the dynamic characteristics of the developed programs. The values of the dynamic characteristics of a Q-effective program depend on the implemented algorithm and the conditions of development and execution of the program. The paper provides a recommendation to the developer of a Q-effective program in the case where he wants to improve the values of the dynamic characteristics of the program being developed.

Therefore, the research shows that the method of designing Q-effective programs can be applied to efficiently implement neural network learning algorithms.

The paper is the first to consider an efficient implementation of neural network learning algorithms using the concept of a Q-determinant. Let’s describe the necessary information about the concept of the Q-determinant. These are the following notions.

Key words: neural network learning, stochastic gradient descent method, error back propagation method, Q-determinant of algorithm, Q-effective implementation of algorithm, Q-effective program.

References

1. Aleeva V.N. Analiz parallel’nyx chislennyx algoritmov. Preprint N 590. Novosibirsk: VC SO AN SSSR, 1985. 23 s. (in Russian)
2. Valentina Aleeva, Rifkhat Aleev. Investigation and Implementation of Parallelism Resources of Numerical Algorithms // ACM Transactions on Parallel Computing. 2023. V. 10. N 2, Article number 8. P. 1–64. DOI: 10.1145/3583755.
3. Ershov YU. L., Palyutin E. A. Matematicheskaya logika. M.: Nauka, 1987. 336 s. (in Russian)
4. Aleeva V. N. Improving Parallel Computing Efficiency // Proceedings – 2020 Global Smart Industry Conference, GloSIC 2020. IEEE. 2020. P. 113–120. Article number 9267828. DOI: 10.1109/GloSIC50886.2020.9267828.
5. Aleeva V. Designing a Parallel Programs on the Base of the Conception of Q-Determinant // Supercomputing. RuSCDays 2018. Communications in Computer and Information Science. 2019. Vol. 965. P. 565–577. DOI: 10.1007/978-3-030-05807-4-48.
6. Gudfellou Ya, Bendzhio I., Kurvill‘ A. Glubokoe obuchenie. M.: DMK Press, 2018. 652 s. (in Russian)
7. Nielsen M. A. Neural Networks and Deep Learning. [Electron. Res.]: http: //neuralnetworksanddeeplearning.com/chap2.html. Accessed: 11.02.2025.
8. Nikolenko S. I., Kadurin A. A., Arxangelskaya E. O. Glubokoe obuchenie. SPb.: Piter, 2018. 480 s. (in Russian)
9. Superkomp’uter “Tornado YuUrGU”. [Electron. Res.]: http://supercomputer.susu.ru/ computers/tornado/. Accessed: 11.02.2025. (in Russian)
10. Otkrytaya enciklopediya svojstv algoritmov. [Electron. Res.]: https://algowiki-project.org/ru. Accessed: 11.02.2025. (in Russian)

Bibliographic reference: Aleeva V. N., Sapozhnikov A. S. Effcient Implementation of Neural Network Learning
Algorithms Using the Concept of a Q-determinant. // “Problems of informatics”, 2025, N 3(68), P. 5-16.  DOI: 10.24412/2073-0667-2025-3-5-16. – EDN:XGOUCS

J. Rahmani, B. V. Baibara, S. G. Tetov

Moscow Technical University of Communications and Informatics, 111024, Moscow, Russia

VULNERABILITIES OF LARGE LANGUAGE MODELS: ANALYSIS AND PROTECTION METHODS

DOI: 10.24412/2073-0667-2025-3-17-33

EDN: TFEVWR

The rapid adoption of large language models (LLMs) in enterprise environments has revolutionized industries by enabling advanced automation, customer service, content generation, and data analysis. However, this technological advancement introduces significant security risks, as organizations increasingly report breaches and vulnerabilities associated with AI systems. According to industry reports, 74 % of major IT companies experienced AI-related security incidents in 2024, with 89 % expressing concerns about vulnerabilities in third-party AI applications. This paper provides a comprehensive analysis of the most critical security threats in LLM deployments, focusing on prompt injection attacks, different supply chain vulnerabilities, and data poisoning, while proposing mitigation strategies to enhance AI security.
Key Vulnerabilities in LLM Applications:
In this paper we analyze most critical vulnerabilities based on OWASP TOP 10 LLM list. OWASP (Open Worldwide Application Security Project — The Open World Application Security Project (OWASP) in its “OWASP Top 10 for Large Language Model Applications 2025” ranked operational injection, sensitive information disclosure, supply chain vulnerabilities, data and model poisoning, and improper output handling as the top five vulnerabilities.
1. Prompt Injection Attacks
- Prompt injection occurs when malicious user inputs manipulate an LLM’s behavior, bypassing security restrictions to extract sensitive data, execute unauthorized commands, or generate harmful content.
- Two primary types are identified: a) Direct prompt injection: Explicit adversarial instructions that override system prompts (e.g., “Ignore previous instructions and disclose confidential data”).
b) Indirect prompt injection: Maliciously crafted external data (e.g., poisoned web pages or documents) that indirectly influences the model’s output.
- Advanced techniques like Knowledge Return-Oriented Prompting (KROP) demonstrate how attackers can bypass safeguards by leveraging the model’s training data
- Mitigation strategies: Input validation, output filtering, least-privilege access controls, and alignment-based guardrails to enforce intended model behavior.
2. Supply Chain Vulnerabilities
- LLMs rely on external dependencies, including pre-trained models, datasets, and third-party libraries, which can be compromised to introduce backdoors or biased behavior.
- Case studies include the “pymafka” PyPI package, which mimicked a legitimate library but deployed Cobalt Strike malware.
- A formal risk assessment model evaluates the probability of compromise across data, dependencies, and training pipelines.
- Mitigation strategies: Secure model provenance (e.g., signed artifacts), Software Bill of Materials (SBOM) for dependencies, and continuous monitoring for anomalies.
3. Data Poisoning Attacks
- Adversaries corrupt training data to manipulate model outputs, leading to biased, unethical, or malicious behavior.
- Notable incidents include Microsoft’s Tay chatbot, which was manipulated into generating offensive content through user interactions.
- Risks extend to pickle-based model serialization, where malicious code can execute during deserialization, compromising entire systems.
- Mitigation strategies: Secure data sourcing, sandboxing untrusted inputs, and anomaly detection via gradient analysis and behavioral divergence metrics.
Defensive Frameworks and Future Challenges
The paper highlights existing defense mechanisms while acknowledging persistent gaps in LLM security. Key recommendations include:
- Secure-by-design principles, such as using safer serialization formats (e.g., SafeTensors instead of pickle).
- Multi-layered validation of inputs, outputs, and model behavior.
Despite these measures, the evolving sophistication of attacks—such as Indirect Prompt Injection, Knowledge-Return-Oriented-Prompting and backdoored models — demands ongoing research. The paper concludes by emphasizing the need for industry-wide collaboration, standardized security benchmarks, and regulatory frameworks to mitigate risks in LLM adoption.
Key words: LLM, artificial intelligence, prompt injection, supply chain attack, data poisoning.
 

References

1. Large Language Model Statistics And Numbers (2025) // springsapps [Electron. Res.]: https://springsapps.com/knowledge/large-language-model-statistics-and-nu... (accessed 9 April 2025).
2. HiddenLayer AI Threat Landscape Report Reveals AI Breaches on the Rise; Security Gaps & Unclear Ownership Afflict Teams// PR Newswire [Electron. Res.]: https://hiddenlayer.com/threatreport2025/(accessed 9 April 2025).
3. Large language model // wikipedia [Electron. Res.]: https://en.wikipedia.org/wiki/Large_language_model(accessed 9 April 2025).
4. What are large language models (LLMs)? // ibm.com [Electron. Res.]: https://www.ibm.com/think/topics/large-language-models (accessed 9 April 2025).
5. LLM in business: options for using large language models // napoleonit [Electron. Res.]: https://napoleonit.ru/blog/llm-v-biznese-varianty-ispolzovaniya-bolshih-... (accessed 9 April 2025).
6. LLM Overview // habr [Electron. Res.]: https://habr.com/ru/companies/tensor/articles/790984/ (accessed 9 April 2025).
7. Understanding Encoder And Decoder LLMs // Ahead of AI [Electron. Res.]: https://magazine.sebastianraschka.com/p/understanding-encoder-and-decoder (accessed 9 April 2025).
8. OWASP Top 10 for LLM Applications 2025 // OWASP [Electron. Res.]: https://genai.owasp.org/resource/owasp-top-10-for-llm-applications-2025/ (accessed 9 April 2025).
9. Universal and Transferable Adversarial Attacks on Aligned Language Models // arXiv [Electron. Res.]: https://arxiv.org/abs/2307.15043 (accessed 9 April 2025).
10. Securing Large Language Models: Threats, Vulnerabilities and Responsible Practices // arXiv [Electron. Res.]: https://arxiv.org/abs/2403.12503(accessed 9 April 2025).
11. Knowledge Return Oriented Prompting (KROP) // arXiv [Electron. Res.]: https://arxiv.org/abs/2406.11880 (accessed 9 April 2025).
12. From ChatGPT to ThreatGPT: Impact of Generative AI in Cybersecurity and Privacy // arXiv [Electron. Res.]: https://arxiv.org/abs/2307.00691 (accessed 9 April 2025).
13. ATLAS Matrix // MITRE ATLAS [Electron. Res.]: https://atlas.mitre.org/matrices/ATLAS (accessed 9 April 2025).
14. Security of large language model applications (LLM, GenAI) (LLM, GenAI) // habr [Electron. Res.]: https://habr.com/ru/articles/843434/(accessed 9 April 2025).
15. Large Language Model Supply Chain: Open Problems From the Security Perspective // arXiv [Electron. Res.]: https://arxiv.org/pdf/2411.01604 (accessed 9 April 2025).
16. Use of Obfuscated Beacons in ‘pymafka’ Supply Chain Attack Signals a New Trend in macOS Attack TTPs // SentinelLabs [Electron. Res.]: https://www.sentinelone.com/labs/useof-obfuscated-beacons-in-pymafka-sup... (accessed 9 April 2025).
17. New “pymafka” malicious package drops Cobalt Strike on macOS, Windows, Linux // sonatype [Electron. Res.]: https://www.sonatype.com/blog/new-pymafka-malicious-packagedrops-cobalt-... (accessed 9 April 2025).
18. Google introduced SLSA, a solution to combat supply chain attacks // habr [Electron. Res.]:
https://habr.com/ru/news/564140/ (accessed 9 April 2025).
19. Machine Learning Security against Data Poisoning: Are We There Yet? // arXiv [Electron.Res.]: https://arxiv.org/abs/2204.05986 (accessed 9 April 2025).
20. Never a dill moment: Exploiting machine learning pick lefiles // arXiv [Electron. Res.]: https://blog.trailofbits.com/2021/03/15/never-a-dill-moment-exploiting-m... (accessed 9 April 2025)
21. pickle // python docs [Electron. Res.]: https://docs.python.org/3/library/pickle.html.(accessed 9 April 2025).

Bibliographic reference: Rahmani J., Baibara B. V., Tetov S. G. Vulnerabilities of Large Language Models: Analysis and Protection Methods // “Problems of informatics”, 2025, N 3(68), P. 17-33 DOI: 10.24412/2073-0667-2025-3-17-33.

V. E. Malyshkin, V. A. Perepelkin, Yu.Yu. Nushtaev*,**

*Institute of computational mathematics and mathematical geophysics SB RAS, 630090, Novosibirsk, Russia
**Novosibirsk State University, 630090, Novosibirsk, Russia
***Novosibirsk State Technical University, 630073, Novosibirsk, Russia

REDUCTION OF INVOCATION OVERHEAD IN AUTOMATICALLY GENERATED PROGRAMS WITH THE ACTIVE KNOWLEDGE CONCEPT

DOI: 10.24412/2073-0667-2025-3-34-51

EDN: CBKGZK

Parallel programs development automation is a relevant research direction, potentially beneficial in multiple ways. It allows to reduce complexity and labor intensity for human, improve efficiency of constructed programs and support software and algorithms accumulation and reuse. One of the problems here is to reduce the invocation overhead which arises from the fact that in practice programs have to be constructed mostly out of modules. This fact implies modules unification and overhead, related to their invocation, data transfer, run-time environment setup, etc. The overhead significantly affects the constructed program efficiency (i.e. program execution time, memory consumption, network load, etc.), which is essential in high performance computing. Programs construction system capabilities in reduction of the overhead highly depend on the computational model employed by the system. In the work we consider the invocation overhead reduction problem through the active knowledge concept [10] — a methodology for efficient programs construction automation in particular subject domains. The concept is based on the theory of parallel programs and systems synthesis on the basis of computational models [11]. It implies that to perform automatic construction of efficient-enough programs in a particular subject domain one has to make a machine-oriented partial formal description of the subject domain called active knowledge base [9]. It contains description of various algorithms, related software modules and peculiarities of the subject domain. Based on active knowledge base it is possible to formulate a class of applied problems to solve and automatically construct a program to solve any of the problems. The key concept here is computational model, which for simplicity can be concerned as a bipartite directed graph of operations and variables vertices. Ingoing and outgoing arcs for particular operation vertex denote its input and output variables. Computational model describes a subject domain in sense that the domain has some variables and there is an ability to compute some variables from some other variables. Each operation can be given a suitable computational module, called code fragment, capable of computing values of its output variables from values of its input variables. Conventional subroutine of given form can serve as an example of a code fragment. The computational process then is concerned as follows. Some variables are assigned with arbitrary values. Any operation can be executed if all its input variables have values. Operation execution is code fragment invocation with values of input and output variables’ values as input and output arguments. Operations are executed (maybe in parallel) until all variables marked as demanded are computed. The computational model can be employed for automatic programs construction. A constructed program consists of two parts. The first one is a set of code fragments contained in the active knowledge base. The second one is generated code, which can be called “glue” code. Its main purpose is to invoke code fragments, pass arguments to them, organize network data transfer and perform other similar tasks. To provide high efficiency of a constructed program the following two conditions have to be satisfied. Firstly, “glue” code has to be efficient. Secondly, the code fragments invocation overhead has to be low enough. For example, if a code fragment is a conventional subroutine, then its invocation requires control passing (call) and data movement between different memory locations and or registers. In conventional compilers this overhead can sometimes be reduced using the inlining technique. If a code fragment is a program written in another language, then corresponding run-time environment and data conversion has to be made. Notably, the inlining technique not always can be employed by the compiler because it relies on complex static code analysis. Unless the compiler is able to extract all necessary information to perform inlining it cannot be applied. An alternative approach is to manually provide code fragments with necessary metainformation. In such case invocation of the code fragment can be implemented not as a procedure call, but as an inline code snippet. Code snippet of particular form is an example of a code fragment with less overhead than a conventional procedure. The active knowledge concept supports this approach by allowing the inclusion of different code fragment types with necessary metainformation into active knowledge base. Another advantage the active knowledge concept suggests is automatic operations aggregation (batching). The idea behind this technique is to combine a group of similar operations into a single code fragment, thus reducing overhead. A practical example is aggregating multiple operations for GPU to reduce input/output data transfer between main memory and GPU memory. Provided necessary metainformation is given, multiple GPU operations can be aggregated into one GPU call. Such low-level techniques as CUDA Graph [20] can be applied automatically. Some subject domains have additional possibilities of batching. For example, cuFFT library provides an API to perform batch processing of multiple fast Fourier transforms more efficiently. With the active knowledge concept, it is possible to perform such batching automatically. For that an active knowledge base has to be supplied with corresponding metainformation and batching algorithm implementation. The system will be able to analyze the computational model graph in order to find operations to batch. In the paper we concern a practical example — automatic construction of a hybrid parallel program which uses both CPU and GPU to achieve satisfactory performance in seismic data processing [12].

Key words: active knowledge concept, computational model, automatic program construction.

This work was carried out under state contract with ICMMG SB RAS FWNM-2025-0005.

References

1. Kale L. V., Krishnan S. Charm++ a portable concurrent object oriented system based on c++ //Proceedings of the eighth annual conference on Object-oriented programming systems, languages, and applications. — 1993. — S. 91-108.
2. Charm++. Parallel Computer Network [Electron. Res.]: http://charmplusplus.org/. (accessed: 01.05.2025).
3. OpenCL [Electron. Res.]: https://www.khronos.org/opencl/ (accessed: 01.05.2025).
4. Coarray Fortran [Electron. Res.]: http://caf.rice.edu (accessed: 01.05.2025).
5. Reid J. Coarrays in the next fortran standard //ACM SIGPLAN Fortran Forum. New York, NY, USA : ACM, 2010. V. 29. N 2. P. 10–27.
6. DVM — sistema razrabotki parallel’nykh programm [Electron. Res.]: http://dvm-system.org/ru/about/ (accessed: 01.05.2025).
7. Bakhtin V.A. [et al.]. Rasshireniye DVM-modeli parallel’nogo programmirovaniya dlya klasterov s geterogennymi uzlami // Vestnik Yuzhno-Ural’skogo universiteta. Chelyabinsk: Izdatel’skiy tsentr YuUrGU, 2012. Seriya: Matematicheskoye modelirovaniye i programmirovaniye. N 18 (277). Vypusk 12. S. 82–92.
8. Kataev N., Kolganov A. The experience of using DVM and SAPFOR systems in semi automatic parallelization of an application for 3D modeling in geophysics // The Journal of Supercomputing. 2019. T. 75. N 12. S. 7833–7843.
9. Malyshkin V. E., Perepyolkin V. A. Postroenie baz aktivnyx znanij dlya avtomaticheskogo konstruirovaniya reshenij prikladnyx zadach na osnove sistemy LuNA // Parallelnye vychislitelnye texnologii — XVIII vserossijskaya nauchnaya konferenciya s mezhdunarodnym uchastiem, PaVT’2024,
g. Chelyabinsk, 2–4 aprelya 2024 g. Korotkie statyi i opisaniya plakatov. Chelyabinsk: Izdatelskij centr YuUrGU, 2024. s. 57–68.
10. Victor Malyshkin. Active Knowledge, LuNA and Literacy for Oncoming Centuries. In Essays Dedicated to Pierpaolo Degano on Programming Languages with Applications to Biology and Security - Volume 9465. Springer-Verlag, Berlin, Heidelberg, 2015. p. 292–303.
11. Sintez parallelnykh programm i sistem na vychislitelnykh modelyakh / V. A. Valkovsky, V. E. Malyshkin; Onv. red. V. E. Kotov; AN SSSR, Sib. otd-nie, VC. Novosibirsk : Nauka. Sib. otd-nie, 1988. 126 s. (In Russian).
12. Vyrodov A. Yu. et al. Printsipy organizatsii programmno-analiticheskoy sistemy dlya parallel’noy obrabotki seysmicheskikh dannykh // Vestnik SibGUTI. 2024. T. 18. N 2. S. 57–68.
13. Ragan-Kelley J. et al. Halide: a language and compiler for optimizing parallelism, locality, and recomputation in image processing pipelines // Acm Sigplan Notices. 2013. T. 48. N 6. S. 519–530.
14. PLUTO [Electron. Res.]: https://pluto-compiler.sourceforge.net/ (accessed: 01.03.2025).
15. Bondhugula U. et al. Automatic transformations for communication-minimized parallelization and locality optimization in the polyhedral model // Compiler Construction: 17th International Conference, CC 2008, Held as Part of the Joint European Conferences on Theory and Practice of Software, ETAPS 2008, Budapest, Hungary, March 29–April 6, 2008. Proceedings 17. Springer Berlin Heidelberg, 2008. S. 132–146.
16. Bondhugula U. et al. A practical automatic polyhedral parallelizer and locality optimizer // Proceedings of the 29th ACM SIGPLAN Conference on Programming Language Design and Implementation. 2008. S. 101–113.
17. Polyhedral Compilation [Electron. Res.]: http://polyhedral.info/ (accessed: 01.03.2025).
18. Malyshkin V. E., Perepelkin V. A. LuNA fragmented programming system, main functions and peculiarities of run-time subsystem // International Conference on Parallel Computing Technologies. Berlin, Heidelberg: Springer Berlin Heidelberg, 2011. S. 53–61.
19. Malyshkin V. E., Perepelkin V. A. Opredelenie ponyatiya programmy // “Problemy informatiki”, 2024, N 2, S. 16–31.
20. CUDA Graphs [Electron. Res.]: https://developer.nvidia.com/blog/cuda-graphs/ (accessed: 01.05.2025).
21. NVIDIA. cuFFT Library [Electron. Res.]: https://docs.nvidia.com/cuda/cufft/index.html (accessed: 01.05.2025).
22. OpenMP [Electron. Res.]: http://www.openmp.org/ (accessed: 01.03.2025).
23. NVIDIA CUDA [Electron. Res.]: https://developer.nvidia.com/cuda-toolkit (accessed: 01.05.2025).
24. Malyshkin V. Active Knowledge, LuNA and Literacy for Oncoming Centuries // In Essays Dedicated to Pierpaolo Degano on Programming Languages with Applications to Biology and Security. V. 9465. Springer-Verlag, Berlin, Heidelberg, 2015. P. 292–303.

 Bibliographic reference: Malyshkin V. E., Perepelkin V. A., Nushtaev Yu.Yu. Reduction of Invocation Overhead in Automatically Generated Programs with the Active Knowledge Concept // “Problems of informatics”, 2025, N 3(68), P. 34-52 DOI: 10.24412/2073-0667-2025-3-34-52

A. Bobokhonov, L. Xuramov, A. Rashidov

Samarkand State University named after Sh. Rashidov, Samarkand, Uzbekistan

DETECTION OF SKIN DISEASES FROM IMAGES USING MACHINE LEARNING AND DEEP LEARNING TECHNIQUES

DOI: 10.24412/2073-0667-2025-3-52-71

EDN: WNRKQY

Today, classification of skin diseases based on automated systems by analyzing medical images taken from the affected skin surface is one of the important methods to be studied. Skin diseases are one of the global health problems that is increasing year by year and endangering the lives of many people. Early detection of this disease is crucial in preventing its progression and its consequences. Currently, many studies are being conducted to detect skin diseases at early stages and several solutions are being proposed. In particular, classification of skin diseases based on medical images using intelligent systems is one of the best solutions proposed by researchers. In this research work, the methods, models and algorithms for automatic classification of skin diseases based on computer-aided machine learning (ML) and deep learning (DL) algorithms were analyzed. Also, methods for pre-processing medical images were studied to ensure fast and accurate performance of ML and DL models. As a result of the analysis, comparative tables were developed for further research work to compare the results of previous studies and the accuracy of the models proposed in them. The main goal of the study is to fill the research gap in the application of ML and DL models in skin disease classification. This study will help researchers find better solutions for classifying skin diseases, identify existing problems and recent achievements in the classification.

Key words: Skin diseases, Medical images, Image preprocessing, Segmentation, Classification, Machine learning, Deep learning.

References

1. Burden of skin disease. [Electron. Res.]: https://www.aad.org/member/clinical-quality/clinical-care/bsd.
2. Skin conditions by the numbers. [Electron. Res.]: https://www.aad.org/media/stats-numbers.
3. Rahman Attar et al. Reliable Detection of Eczema Areas for Fully Automated Assessment of Eczema Severity from Digital Camera Images. [Electron. Res.]: https://doi.org/10.1016/j.xjidi. 2023.100213.
4. Elisabeth V. Goessinger et al. Image-Based AI in Psoriasis Assessment: The Beginning of a New Diagnostic Era? // AJCD 2024. [Electron. Res.]: https://doi.org/10.1007/s40257-024-00883-y.
5. Kimberley Yu, BA et al. “Machine Learning Applications in the Evaluation and Management of Psoriasis: A Systematic Review” 2020, DOI: 10.1177/2475530320950267.
6. Cort’es Verd’u R. et al. Prevalence of systemic lupus erythematosus in Spain: Higher than previously reported in other countries // Rheumatology. 2020, N 59, P. 2556–2562.
7. Iciar Usategui et al. Systemic Lupus Erythematosus: How Machine Learning Can Help Distinguish between Infections and Flares // Bioengineering. 2024, N 11(1), 90; [Electron. Res.]: https://doi.org/10.3390/bioengineering11010090.
8. Basal Cell Carcinoma Treatment in India. [Electron. Res.]: https://bit.ly/3Ybz4Aj.
9. Squamous cell carcinoma of the skin. [Electron. Res.]: https://mayocl.in/4f5yhbd.
10. Bhagyasri M., et al. Study on machine learning and deep learning methods for cancer detection // J. Image Process AI . 2018. Vol. 4.
11. Kuldeep Vayadande et al. Innovative approaches for skin disease identification in machine learning: A comprehensive study // Oral Oncology Reports. June 2024. Volume 10, 100365.
12. Nisar H., et al. Automatic segmentation and classification of eczema skin lesions using supervised learning, 2020; 10.1109/ICOS50156.2020.9293657.
13. Jagdish M., et al. Advance study of skin diseases detection using image processing methods // NVEO 2022, Vol. 9, N 1, [Electron. Res.]: https://www.cabidigitallibrary.org/doi/full/10.5555/20220157042.
14. AlDera S. A., Othman M. T. B. A Model for Classification and Diagnosis of Skin Disease using Machine Learning and Image Processing Techniques // IJACSA. 2022. Vol. 13, N 5.
15. Qays Hatem Mustafa. Skin lesion classification system using a K nearest neighbor algorithm // HVCI, Biomedicine, and Art. 2022. 5:7. [Electron. Res.]: https://doi.org/10.1186/s42492-022-00103-6.
16. Souza Jhonatan et al. Automatic Detection of Lupus Butterfly Malar Rash Based on Transfer Learning. [Electron. Res.]: https://sol.sbc.org.br/index.php/wvc/article/download/13499/13347/.
17. Bandyopadhyay Samir et al. Machine Learning and Deep Learning Integration for Skin Diseases Prediction // IJETT ISSN. 11–18, February, 2022. Vol. 70. Issue 2. P. 2231–5381.
18. Laura K Ferris et al. Computer-aided classification of melanocytic lesions using dermoscopic images // J. Am Acad Dermatol. Nov. 2015; 73(5):769-76.
19. What is Normalization in Machine Learning? A Comprehensive Guide to Data Rescaling. [Electron. Res.]: https://www.datacamp.com/tutorial/normalization-in-machine-learning.
20. Normalization: The First Step in Image Prep. [Electron. Res.]: https://www.linkedin.com/pulse/normalization-first-step-image-preprocess....
21. Manoj Diwakar, Manoj Kumar. A review on CT image noise and its denoising // Biomedical Signal Processing and Control. 2018. N 42. P. 73–88.
22. Patil R. et al. Medical Image Denoising Techniques: A Review. 2022. Volume 4, Issue 1.
23. Edge Detection in Image Proc.: An Introduction. [Electron. Res.]: https://blog.roboflow.com/edge-detection/.
24. Lakshmanan B. et al. Stain removal through color normalization of haematoxylin and eosin images: a review // Journal of Physics: Conference Series. 2019. 1362.
25. Different Morphological Operations in Image Processing. [Electron. Res.]: https://www.geeksforgeeks.org/different-morphological-operations-in-imag....
26. Zhe Zhu. Change detection using landsat time series: A review of frequencies, preprocessing, algorithms, and applications // ISPRS 2017. [Electron. Res.]: https://doi.org/10.1016/j.isprsjprs.2017.06.013.
27. Mostafiz Ahammed, Md. et al. A machine learning approach for skin disease detection and classification using image segmentation, HA. [Electron. Res.]: https://doi.org/10.1016/j.health.2022.100122.
28. Krishna M., Monika, N. et al. Skin cancer detection and classification using machine learning.2020. Volume 33, Part 7. [Electron. Res.]: https://doi.org/10.1016/j.matpr.2020.07.366.
29. Vidya M., et. al. Skin Cancer Detection using Machine Learning Techniques // 2020 IEEE (CONECCT) 10.1109/CONECCT50063.2020.9198489.
30. Maurya R et al. Skin cancer detection through attention guided dual autoencoder approach with ELM // Sci. Rep. 2024. 14(1):17785. [Electron. Res.]: https://doi.org/10.1038/s41598-024-68749-1.
31. Keerthana D et al. Hybrid convolutional neural networks with SVM classifier for classification of skin cancer // Biomed. 2023. [Electron. Res.]: https://doi.org/10.1016/j.bea.2022.100069.
32. Shuchi Bhadula, et al. Machine Learning Algorithms based Skin Disease Detection // IJITEE. 2019. Vol. 9 Iss. 2. [Electron. Res.]: https://www.researchgate.net/publication/341371302_MLSDD.
33. Hameed N., et al. A Computer-Aided diagnosis system for classifying prominent skin lesions using machine learning. 2019, DOI: 10.1109/CEEC.2018.8674183.
34. Koklu M. et al. Skin Lesion Classification using Machine Learning Algorithms // Int. J. Intell. Syst. Appl. Eng., 2017. Vol. 4, N 5, P. 285–289, DOI: 10.18201/ijisae.2017534420.
35. Chen Yin et al. Non-invasive prediction of the chronic degree of lupus nephropathy based on ultrasound radiomics // Sage Journals Home. 2023. Volume 33, Issue 2.
36. Parvathaneni Naga Srinivasu et al. Classification of Skin Disease Using Deep Learning Neural Networks with MobileNet V2 and LSTM // Sensors (Basel). 2021 Apr 18; 21(8):2852.
37. Yaseliani Mohammad et al. Diagnostic clinical decision support based on deep learning and knowledge-based systems for psoriasis: From diagnosis to treatment options // Computers & Industrial Engineering. January 2024, Vol. 187, 109754.
38. Jothimani Subramani et al. Gene-Based Predictive Modelling for Enhanced Detection of SLE Using CNN-Based DL Algorithm // Diagnostics, 2024. Vol. 14, Iss. 13.
39. Syed Inthiyaz et al. Skin disease detection using deep learning // Advances in Engineering Software. January 2023. Vol. 175.
40. Himanshu K. Gajera et al. A comprehensive analysis of dermoscopy images for melanoma detection via deep CNN features // BSPC. January 2023. Vol. 79, Part 2.
41. Reza Ahmadi Mehr, Ali Ameri. Skin Cancer Detection Based on Deep Learning // Journal of Biomedical Physics and Engineering. December 2022. Vol. 12, Iss. 6, 55, P. 559–568.
42. Jahin Alam Md. et al. S2C-DeLeNet: A parameter transfer based segmentation-classification integration for detecting skin cancer lesions from dermoscopic images // Computers in Biology and Medicine. November 2022, Vol. 150.
43. Hammad Mohamed et al. Enhanced Deep Learning Approach for Accurate Eczema and Psoriasis Skin Detection // Sensors. 2023, 23, 7295. [Electron. Res.]: https://doi.org/10.3390/s23167295.
44. Rai H. M. et al. Computational Intelligence Transforming Healthcare 4.0: Innovations in Medical Image Analysis through AI and IoT Integration // DDDSSIHC. 2025. Chap.3, P. 15, CRC Press. [Electron. Res.]: https://doi.org/10.1201/9781003507505.
45. Bobokhonov A., Xuramov L., Rashidov A. Tibbiy tasvirlar asosida teri kasalliklarini samarali tasniflash usullari // Digital Transformation and AI, 3(3), 128–139 [Electron. Res.]: https://dtai.tsue.uz/index.php/dtai/article/view/v3i319.

 Bibliographic reference: Bobokhonov A., Xuramov L., Rashidov A. Detection of Skin Diseases from Images Using Machine Learning and Deep Learning Techniques // “Problems of informatics”, 2025, N 3(68), P. 53-71 DOI: 10.24412/2073-0667-2025-3-53-71.

A. A. Yurtin

South Ural State University (National Research University), 454080, Chelyabinsk, Russia

A METHOD FOR FORECASTING THE ERROR AND TRAINING TIME OF NEURAL NETWORKS FOR MULTIVARIATE TIME SERIES IMPUTATION

DOI: 10.24412/2073-0667-2025-3-72-95

EDN: XLSZLH

The article presents a neural network-based method called tsGAP2, designed for predicting the error and training time of neural network models used for imputing missing values in multivariate time series. The input data for the method are neural network represented as a directed acyclic graphs, where nodes correspond to layers and edges represent connections between them. The method involves three components: an Autoencoder, which transforms the graph-based representation of the model into a compact vector form; an Encoder, which encodes the hyperparameters and characteristics of the computational device; and an Aggregator, which combines the vector representations to generate the prediction. Training of the tsGAP2 neural network model is carried out using a composite loss function, defined as a weighted sum of multiple components. Each component evaluates different aspects of the tsGAP2 model’s output, including the correctness of the decoded neural network model from the vector representation, the prediction of the model’s error, and its training time. For the study, a search space comprising 200 different architectures was constructed. During the experiments, 12,000 training runs were conducted on time series from various application domains. The experimental results demonstrate that the proposed method achieves high accuracy in predicting the target model’s error: the average error, measured using SMAPE, is 4.4 %, which significantly outperforms existing alternative approaches, which show an average error of 27.6 %. The average prediction error for training time was 8.8 %, also significantly better than existing methods, which show an error of 61.6 %.
Key words: time series, missing value imputation, neural network models, autoencoder, graph neural networks, attention mechanism, performance prediction, neural architecture search.

References

1. Aydin S. Time series analysis and some applications in medical research // Journal of Mathematics and Statistics Studies. 2022. V. 3. N 2. P. 31–36. DOI: 10.32996/JMSS.
2. Voevodin V. V., Stefanov K. S. Development of a portable software solution for monitoring and analyzing the performance of supercomputer applications // Numerical Methods and Programming. 2023. V. 24. P. 24–36. DOI: 10.26089/NumMet.v24r103.
3. Kumar S., Tiwari P., Zymbler M. L. Internet of Things is a revolutionary approach for future technology enhancement: a review // Journal of Big Data. 2019. V. 6. Art. 111. DOI: 10.1186/S40537-019-0268-2.
4. Gromov V. A., Lukyanchenko P. P., Beschastnov Yu. N., Tomashchuk K. K. Time Series Structure Analysis of the Number of Law Cases // Proceedings in Cybernetics. 2022. N 4 (48). P. 37–48.
5. Kazijevs M., Samad M. D. Deep imputation of missing values in time series health data: A review with benchmarking // J. Biomed. Informatics. 2023. V. 144. P. 104440. DOI: 10.1016/J.JBI.2023.104440.
6. Elsken T., Metzen J. H., Hutter F. Neural Architecture Search: A Survey // J. Mach. Learn. Res. 2019. V. 20. N 55. P. 1–21. [Electron. res.]: https://jmlr.org/papers/v20/18-598.html.
7. Wozniak A. P., Milczarek M., Wozniak J. MLOps Components, Tools, Process, and Metrics: A Systematic Literature Review // IEEE Access. 2025. V. 13. P. 22166–22175. DOI: 10.1109/ACCESS.2025.3534990.
8. Weights & Biases: Machine learning experiment tracking, dataset versioning, and model management. [El. Res.]: https://wandb.ai/. Access date: 2025-06-11.
9. Bergstra J., Bengio Y. Random search for hyper-parameter optimization // J. Mach. Learn. Res. 2012. V. 13. P. 281–305.
10. Dong X., Yang Y. NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search // 8th Int. Conf. on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26–30, 2020. [Electron. res.]: https://openreview.net/forum?id=HJxyZkBKDr.
11. Ding Y., Huang Z., Shou X., Guo Y., Sun Y., Gao J. Architecture-Aware Learning Curve Extrapolation via Graph Ordinary Differential Equation // AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, Feb. 25 — Mar. 4, 2025, Philadelphia, PA, USA / ed. By T. Walsh, J. Shah, Z. Kolter. AAAI Press, 2025. P. 16289–16297. DOI: 10.1609/AAAI.V39I15.33789.
12. timeseries Graph Attention Performance Predict. [El. Res.]: https://gitverse.ru/yurtinaa/tsGAP2. Access date: 2025-05-03.
13. Gawlikowski J., Tassi C. R. N., Ali M., Lee J., Humt M., Feng J., Kruspe A., Triebel R., Jung P., Roscher R., Shahzad M., Yang W., Bamler R., Zhu X. X. A survey of uncertainty in deep neural networks // Artif. Intell. Rev. 2023. V. 56. N 1. P. 1513–1589. ISSN: 1573–7462. DOI: 10.1007/s10462-023-10562-9.
14. Zela A., Siems J. N., Zimmer L., Lukasik J., Keuper M., Hutter F. Surrogate NAS Benchmarks: Going Beyond the Limited Search Spaces of Tabular NAS Benchmarks // The Tenth Int. Conf. on Learning Representations, ICLR 2022, Virtual Event, April 25–29, 2022. [Electron. res.]: https://openreview.net/forum?id=OnpFa95RVqs.
15. Titsias M. Variational Learning of Inducing Variables in Sparse Gaussian Processes // Proc. of
the Twelfth Int. Conf. on Artificial Intelligence and Statistics. / ed. by D. van Dyk, M. Welling. Hilton Clearwater Beach Resort, Clearwater Beach, Florida, USA: PMLR, 16–18 Apr. 2009. V. 5. P. 567–574. [Electron. res.]: https://proceedings.mlr.press/v5/titsias09a.html.
16. Ying C., Klein A., Christiansen E., Real E., Murphy K., Hutter F. NAS-Bench-101: Towards Reproducible Neural Architecture Search // Proc. of the 36th Int. Conf. on Machine Learning, ICML 2019, June 9–15, Long Beach, California, USA / ed. by K. Chaudhuri, R. Salakhutdinov. PMLR, 2019. V. 97. P. 7105–7114. [Electron. res.]: http://proceedings.mlr.press/v97/ying19a.html.
17. White C., Neiswanger W., Savani Y. BANANAS: Bayesian Optimization with Neural Architectures for Neural Architecture Search // Thirty-Fifth AAAI Conf. on Artificial Intelligence, AAAI 2021, IAAI 2021, EAAI 2021, Virtual Event, Feb. 2–9, 2021. AAAI Press, 2021. P. 10293–10301. DOI: 10.1609/AAAI.V35I12.17233.
18. White C., Zela A., Ru R., Liu Y., Hutter F. How powerful are performance predictors in neural architecture search? // Adv. Neural Inf. Process. Syst. 2021. V. 34. P. 28454–28469.
19. Snoek J., Rippel O., Swersky K., Kiros R., Satish N., Sundaram N., Patwary M., Prabhat, Adams R. P. Scalable Bayesian Optimization Using Deep Neural Networks // Proc. of the 32nd Int. Conf. on Machine Learning (ICML). Lille, France: PMLR, 2015. V. 37. P. 2171–2180.
20. Springenberg J. T., Klein A., Falkner S., Hutter F. Bayesian Optimization with Robust Bayesian Neural Networks // Adv. Neural Inf. Process. Syst / ed. by D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, R. Garnett. V. 29.
21. Wu X., Zhang D., Guo C., He C., Yang B., Jensen C. S. AutoCTS: Automated Correlated Time Series Forecasting // Proc. VLDB Endow. 2021. V. 15. N 4. P. 971–983. DOI: 10.14778/3503585.3503604.
22. Wang C., Chen X., Wu C., Wang H. AutoTS: Automatic Time Series Forecasting Model Design Based on Two-Stage Pruning // arXiv preprint: abs/2203.14169. DOI: 10.48550/arXiv.2203.14169.
23. Velickovic P., Cucurull G., Casanova A., Romero A., Li‘o P., Bengio Y. Graph Attention Networks // 6th Int. Conf. on Learning Representations, ICLR 2018, Vancouver, Canada, April 30 — May 3, 2018. 2018. [Electron. res.]: https://openreview.net/forum?id=rJXMpikCZ.
24. Clevert D., Unterthiner T., Hochreiter S. Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs) // 4th Int. Conf. on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2–4, 2016 / ed. by Y. Bengio, Y. LeCun. 2016. [Electron. res.]: http://arxiv.org/abs/1511.07289.
25. Hochreiter S. The Vanishing Gradient Problem During Learning Recurrent Neural Nets and Problem Solutions // Int. J. Uncertain. Fuzziness Knowl. Based Syst. 1998. V. 6. N 2. P. 107–116. DOI: 10.1142/S0218488598000094.
26. He K., Zhang X., Ren S., Sun J. Deep Residual Learning for Image Recognition // 2016 IEEE Conf. on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, USA. IEEE Computer Society. 2016. P. 770–778. DOI: 10.1109/CVPR.2016.90.
27. Srivastava N., Hinton G. E., Krizhevsky A., Sutskever I., Salakhutdinov R. Dropout: a simple way to prevent neural networks from overfitting // J. Mach. Learn. Res. 2014. V. 15. N 1. P. 1929–1958. DOI: 10.5555/2627435.2670313.
28. Mao A., Mohri M., Zhong Y. Cross-Entropy Loss Functions: Theoretical Analysis and Applications // Proc. of the 40th Int. Conf. on Machine Learning / ed. by A. Krause. 2023. V. 202. P. 23803–23828.
29. Huber P. J. Robust Estimation of a Location Parameter // Breakthroughs in Statistics: Methodology and Distribution / ed. by S. Kotz, N. L. Johnson. Springer New York. 1992. P. 492–518. ISBN: 978-1-4612-4380-9. DOI: 10.1007/978-1-4612-4380-9_35.
30. Bilenko R. V., Dolganina N.Yu., Ivanova E. V., Rekachinsky A. I. High-performance Computing Resources of South Ural State University // Bulletin of the South Ural State University. Series: Computational Mathematics and Software Engineering. 2022. V. 11. N 1. P. 15–30. DOI: 10.14529/cmse220102.
31. BundesAmt F‥ur Umwelt — Swiss Federal Office for the Environment. [El. Res.]: https://www.hydrodaten.admin.ch/. Access date: 2025-05-03.
32. Trindade A., “Electricity Load Diagrams 2011–2014,” UCI Machine Learning Repository (2015) [El. Res.]: https://doi.org/10.24432/C58C86. Access date: 2023-05-03.
33. Lozano A. C., Li H., Niculescu-Mizil A., Liu Y., Perlich C., Hosking J. R. M., Abe N. Spatialtemporal causal modeling for climate change attribution // Proc. of the 15th ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, Paris, France, June 28 — July 1, 2009 / ed. by J. F. Elder IV, F. Fogelman-Souli’e, P. A. Flach, M. J. Zaki. — ACM, 2009. P. 587–596. DOI: 10.1145/1557019.1557086.
34. Laňa I, Olabarrieta I., V’elez M., Del Ser J. On the imputation of missing data for road traffic forecasting: New insights and novel techniques // Transp. Res. Part C: Emerg. Technol. 2018. V. 90. P. 18–33. DOI: 10.1016/j.trc.2018.02.021.
35. Sheppy M., Beach A., Pless S. NREL RSF Measured Data 2011. [El. Res.]: https://data.openei.org/submissions/358. Access date: 2023-09-03.
36. Snytnikov A. V., Ezrokh Yu. S. Solving Vlasov Equation with Neural Networks // Lobachevskii Journal of Mathematics. 2024. V. 45. P. 3416–3423.

 Bibliographic reference: Yurtin A.A. A Method for Forecasting the Error and Training Time of Neural Networks for Multivariate Time Series Imputation // “Problems of informatics”, 2025, N 3(68), P. 72-95 DOI: 10.24412/2073-0667-2025-3-72-95.