Computer Science
Loading...
ISSN 1508-2806
e-ISSN: 2300-7036
Issue Date
2025
Volume
Vol. 26
Number
No. 3
Description
Journal Volume
Computer Science
Vol. 26 (2025)
Projects
Pages
Articles
Optimized lossless audio compression using DCT energy thresholding and machine learning technique
(Wydawnictwa AGH, 2025) Debnath, Asish; Mondal, Uttam Kr.
This paper proposes a novel lossless audio compression technique, utilizing the Discrete Cosine Transform (DCT) coefficient-controlled technique based on energy thresholding, an XOR-based neural network compression model, and a CNN model. Initially, the DCT is applied to the input audio signal to achieve better energy compaction, followed by transforming selected DCT coefficients into a compressed binary stream. Subsequently, this binary stream is passed to two prediction-based optimized models: an XOR model and a CNN model for further compression.The binary stream is divided into two equal pieces, the data and the key. The XOR neural network model processes the data and key to produce an compressed XORed binary stream. Using a proposed CNN architecture, this stream is further compressed with latent space representations to produce compressed audio data. The simulation findings are analyzed using various statistical and robustness measures and compared with existing approaches.
Attention-based multiple-representation method for fingerprint-presentation-attack detection
(Wydawnictwa AGH, 2025) Uttam, Atul Kumar; Agrawal, Rohit; Jalal, Anand Singh
Fingerprint biometrics are one of the most common authentication mechanisms; however, such systems are often compromised by presentation attacks that are made by presentation-attack instruments. Most fingerprint-presentationattack- detection approaches show poor performance due to the large variations in the presentation-attack instruments and the limited feature representation of the input fingerprint. Therefore, this article proposes a hybrid model of shallow and deep features with multiple representations of input fingerprints. To obtain these shallow and deep features, we first enhanced the texture of the input fingerprint through a novel median adaptive local binary pattern filter and an existing binarized statistical image feature. After this, the input fingerprint image and two textured enhanced images are concatenated along with the channel dimension for multiple representations. Finally, an extended ResNeXt architecture with channel and spatial attention (EResNeXt) was used for relevant feature extraction and presentation attack detection. EResNeXt was evaluated in the LivDet-2015 and LivDet-2017 data sets, and ACEs (average classification errors) were obtained at 0.94 and 0.49, respectively.
Developing explainable machine-learning model using augmented concept activation vector
(Wydawnictwa AGH, 2025) Hassanpour, Reza; Oztoprak, Kasim; Netten, Niels; Busker, Tony; Bargh, Mortaza S.; Choenni, Sunil; Kizildag, Beyza; Kilinc, Leyla Sena
Machine-learning models use high-dimensional feature spaces to map their inputs to the corresponding class labels; however, these features often do not have a one-to-one correspondence with the physical concepts that are understandable by humans. This hinders the ability to provide meaningful explanations for the decisions that are made by these models. We propose a method for measuring the correlation between high-level concepts and the decisions that are made by machine-learning models. Our method can isolate the impact of a given high-level concept and accurately measure it quantitatively. Additionally, this study aims to determine the prevalence of frequent patterns in machine-learning models that often occur in imbalanced data sets. We successfully applied the proposed method to fundus images and managed to quantitatively measure the impacts of the radiomic patterns on the model’s decisions.
Modified honey bee algorithm with random selection of virtual machines for dynamic load balancing
(Wydawnictwa AGH, 2025) Sharmah, Daisy; Bora, Kanak Chandra; Khakhlari, Junumoni
Cloud workloads can overwhelm load balancers, leading to inefficiencies and performance issues. To address these challenges, the honey bee load-balancing algorithm is highly effective in enhancing cloud-resource allocation. Inspired by the foraging behavior of honey bees, this algorithm offers a dynamic approach to resource distribution that adapts to changing workloads in real time. This paper delves into the key features and advantages of honey bee load-balancing, focusing on its dynamic resource allocation, overall response time, and data center processing time. Through a comparative study of existing methodologies, we proposed a modified honey bee load-balancing algorithm that incorporated the random selections of virtual machines. Utilizing the CloudAnalyst tool for simulation, we compares traditional and proposed honey bee load-balancing algorithms to evaluate overall response times and data center processing times using user bases, data centers, virtual machine load balancers, time, service broker policies, and regions. The proposed algorithm demonstrates superior performance in these parameters as compared to the traditional approach when using the same metric values for both algorithms.
GDPKG-LLM: integrating gene, disease, and pharmacogenomics knowledge graphs for cognitive neuroscience using large language models
(Wydawnictwa AGH, 2025) Sarabadani, Ali; Fard, Kheirolah Rahsepar; Dalvand, Hamid
Using the structures of large language models (LLMs) in creating knowledge graphs to understand more about the relationship between the entities of cognitive and biological sciences has become a hot research topic. Due to the great knowledge behind the curtain and the deep connections of this research, it is not possible to use the traditional approaches of machine learning and deep learning. In this study,the main goal is to create a comprehensive and integrated knowledge graph(KG) from the combination of three knowledge sources: Gene Ontology (GO), Disease Ontology (DO), and PharmKG. Large Language Models (LLMs) have been used to create this knowledge base. The main purpose of this KG is to understand the relationships between genes, diseases, and drugs. The proposed approach, GDPKG-LLM, has several key steps, including entity matching, similarity analysis, graph alignment, and using GPT-4. GDPKG-LLM was able to extract more than 16,800 nodes and 838,000 edges from these three knowledge bases and provide a rich KG. This graph provides meaningful relationships, making it a valuable resource for future research in personalized medicine and neuroscience. The reviewed evaluation criteria show the superiority of GDPKG-LLM, which strengthens the validity of this model.