Mahmut Kılıçaslan

Human activity recognition (HAR) on wearable and mobile devices requires models that can operate with limited computational resources while remaining reliable under changing real-world sensing conditions. However, most TinyML-based HAR systems follow the same computational path for every input, regardless of whether an activity pattern is familiar or affected by changes in the user, sensor placement, or sensing configuration. This study presents a shift-aware adaptive TinyML framework that adjusts the inference effort according to the prediction confidence and feature-space distribution shift. A lightweight multi-exit architecture allows confident samples to be classified at early stages, while uncertain or shifted samples are routed to deeper exits. A source-derived feature-space shift score complements confidence-based routing by identifying distribution changes that may not be reflected by predictive uncertainty alone. Persistent severe shifts are additionally considered through a guarded label-free test-time adaptation mechanism that does not require new activity annotations or full model retraining. The framework was evaluated on three publicly available HAR benchmark datasets, REALDISP, WISDM, and PAMAP2, covering complementary sensor-displacement, sensing-domain, unseen-user, and body-location shifts. Across the three benchmarks, explicit shift-aware routing produced a stronger adaptive computational response to changing sensing conditions than confidence-only routing. In REALDISP, the computational cost remained nearly unchanged from IDEAL to SELF conditions with confidence-only routing, whereas shift-aware routing increased the normalized computation by 0.0187 (p = 0.0253). In WISDM and PAMAP2, the increase in computation under the distribution shift was approximately 2.88 and 2.85 times greater with shift-aware routing than with confidence-only routing, respectively. The effect on recognition performance, however, depended on the shift scenario. Shift-aware routing significantly improved the target-domain Macro-F1 in the WISDM WATCH-to-PHONE scenario (ΔMacro-F1 = +0.0096, p = 0.0023), whereas no significant differences were observed for REALDISP SELF or PAMAP2 ANKLE. These findings demonstrate that explicit distribution-shift information can help resource-constrained HAR models allocate additional computation when sensing conditions change, while also showing that deeper inference alone may not overcome severe domain mismatch. The proposed framework provides a practical basis for adaptive inference in TinyML-based wearable and edge sensing systems operating under variable real-world conditions.

Digital image segmentation is a critical step in image analysis, pattern recognition, and computer vision applications. In this study, a hybrid method is proposed to improve the accuracy and efficiency of the segmentation process by combining Spiking Neural Network-based edge detection with the Seeded Region Growing algorithm. Initially, the edges of the image are detected by leveraging the biologically inspired structure of Spiking Neural Networks, allowing for more precise localization of object boundaries. The resulting edge map is then used to guide the Seeded Region Growing algorithm. After edge detection, the remaining homogeneous regions are expanded using Seeded Region Growing, completing the segmentation process. The proposed method has been tested on various images, and the results demonstrate superior performance in terms of edge detection accuracy and region consistency.

Bu bölümde RGB renk uzayında renk indirgeme problemi, çok seviyeli histogram tabanlı eşikleme yöntemleri kullanılarak ele alınmıştır. Otsu, Kapur entropi ve histogramın ağırlık merkezi tabanlı eşikleme teknikleri ile her bir renk kanalı için optimum eşik değerleri belirlenmiş ve bu eşik değerleri kullanılarak renk uzayı alt sınıflara ayrılmıştır. Elde edilen her alt sınıfta yer alan piksel değerleri, ilgili sınıfın ortalama yoğunluk değeri ile temsil edilmiştir. Bu yaklaşım sayesinde görüntünün temel karakteristik özellikleri korunurken veri boyutunda önemli ölçüde azalma sağlanmıştır. Çalışmada farklı eşikleme yöntemlerinin renk dağılımı, sınıf içi homojenlik ve görsel kalite üzerindeki etkileri karşılaştırmalı olarak incelenmiştir. Elde edilen bulgular, çok seviyeli eşikleme tabanlı renk indirgeme yöntemlerinin görüntü işleme uygulamalarında etkili bir ön işleme adımı olarak kullanılabileceğini göstermektedir.

Cryptographically strong substitution boxes (S-boxes) is a crucial issue in block cipher cryptographic design. In this paper, we propose the design of a new method to create 16x16 S-boxes by using the Hippopotamus Optimization Algorithm, or HOA and the Logistic Chaotic Map. We introduce a hybrid architecture of the HOA considering dynamic equilibrium among HOA's exploration and exploitation balance with the sensitivity and unpredictability of chaotic systems for this purpose. The chaotic sequences created by the Logistic Map are employed in the first stage to establish a candidate S-box population with diverse structures. Moving on in the optimization stage, the HOA automatically updates the S-box configuration with nonlinearity-based fitness functions. Experimental evaluations are performed against established cryptographic metrics. The results show that our proposed HOA-Logistic-based 16x16 S-boxes perform more satisfactorily than some of the chaotic, or metaheuristic models, when randomness, key sensitivity, and robustness are included. It is quite apparent that the non-linearity criterion value of 109.5 indicates that this method is quite competitive. The large key space obtained by using the Logistic map can be used to improve security against brute-force and statistical attacks. Taken together, the proposed implementation offers an effective and safe means of large-scale S-box generation as is applied today in symmetric cryptosystems.

Background/Objectives: The aim of this study was to classify nailfold videocapillaroscopy (NVC) images obtained from patients with systemic sclerosis (SSc) and healthy individuals using deep learning methods. Methods: Between January and June 2025, 1280 NVC images were recorded from 50 SSc and 30 healthy individuals. The images were classified by two rheumatologists as normal, early, active, and late-stage SSc patterns. After removing 191 unclassifiable and 112 low-quality images, 977 usable images remained. To ensure balanced classes, 245 normal images were excluded. The final dataset was split into training (70%), validation (20%), and test (10%) sets. Six different deep learning models (MobileNetV3Large, ResNet152V2, Xception, VGG-19, InceptionV3, and NASNetLarge) with varying levels of complexity and computational efficiency were selected to compare their performance. Accuracy, precision, recall, F1 Score, and cross-entropy loss were employed as performance metrics. These metrics are commonly used in the literature to evaluate the effectiveness of classification models. Results: Deep learning models achieved an accuracy ranging from 90.6% to 98.9%, a precision of 93.4% to 98.9%, a recall of 90.6% to 98.8%, an F1 score of 92% to 98.9%, and an ROC AUC performance between 99% and 100%. InceptionV3 demonstrated the best performance with an accuracy of 98.95%, a precision of 98.94%, a recall of 98.80%, and an F1 score of 98.88%. In terms of test loss, the lowest values were observed in the InceptionV3 and NasNetLarge models, both with a loss of 0.03. Overall, the ROC AUC values for all models ranged between 98.99% and 100%, indicating competitive performance. Conclusions: The current findings suggest that deep learning methods may be capable of classifying NVC images as accurately as experienced rheumatologists.

This study presents a comprehensive comparative analysis of machine learning, deep learning, and optimization-based hybrid methods for malicious URL detection on the Malicious Phish dataset. For feature selection and model hyperparameter tuning, the Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Harris Hawk Optimizer (HHO) were employed. Both multiclass and binary classification tasks were addressed using classic machine learning algorithms such as LightGBM, XGBoost, and Random Forest, as well as deep learning models including LSTM, CNN, and hybrid CNN+LSTM architectures, with optimization support also integrated into these models. The experimental results reveal that the ELECTRA-based deep learning model achieved outstanding accuracy and F1-scores of up to 99% in both multiclass and binary scenarios. Although optimization-supported hybrid models also improved performance, the language-model-based ELECTRA architecture demonstrated a significant superiority over classical and optimized approaches. The findings indicate that optimization algorithms are effective in feature selection and enhancing model performance, yet next-generation language models clearly set a new benchmark in malicious URL detection.

Image watermarking serves as a crucial technique for protecting copyrights and verifying the ownership of digital images. The watermarking process involves embedding data into images to prevent unauthorized usage of digital content. In this study, an innovative image watermarking method, using Spiking Neural Networks to embed robust and imperceptible watermarks, is proposed. The proposed method benefits from the edge detection capabilities of spiking neural networks to identify optimal regions for watermark locations. By targeting edge detection, this watermarking approach ensures significant resistance to common image processing attacks such as compression, noise addition, and cropping, while maintaining minimal perceptual distortion. The edge image obtained with the spiking neural network approach is divided into 16 × 16 non-overlapping blocks, and edge pixel definitions are made. Moreover, to increase the security level, the watermark image is scrambled with the help of a chaotic substitution box. The scrambled image is placed on the pixels marked as edges in the edge image. Afterward, it is divided into sub-bands by applying a discrete wavelet transform. The watermark image is inserted into the HH (High-High) band with the help of the discrete wavelet transform and the singular value decomposition approach. In the extraction stage, the HH band of the original image is used together with the watermarked image. Comprehensive experiments are conducted to evaluate the proposed technique, revealing its superiority in preserving both image quality and watermark integrity compared to conventional approaches.

Tuberculosis remains a significant health challenge worldwide, affecting a large population. Therefore, accurate diagnosis of this disease is a critical issue. With advancements in computer systems, imaging devices, and rapid progress in machine learning, tuberculosis diagnosis is being increasingly performed through image analysis. This study proposes three segmentation models based on U-Net, V-Net, and Seg-Net architectures to improve tuberculosis detection using the Shenzhen and Montgomery databases. These deep learning-based methods aim to enhance segmentation accuracy by employing advanced preprocessing techniques, attention mechanisms, and non-local blocks. Experimental results indicate that the proposed models outperform traditional approaches, particularly in terms of the Dice coefficient and accuracy values. The models have demonstrated robust performance on popular datasets. As a result, they contribute to more precise and reliable lung region segmentation, which is crucial for the accurate diagnosis of respiratory diseases like tuberculosis. In evaluations using various performance metrics, the proposed U-Net and V-Net models achieved Dice coefficient scores of 96.43% and 96.42%, respectively, proving their competitiveness and effectiveness in medical image analysis. These findings demonstrate that the Dice coefficient values of the proposed U-Net and V-Net models are more effective in tuberculosis segmentation than Seg-Net and other traditional methods.

Forecasts from machine and deep learning models are vital for traders and investors in the global financial markets. Many different forecasting methods rely on technical patterns. In this study, the LSTM model based on candlesticks and financial variables was used to improve trading forecasts of different types. Japanese candlesticks are among the most widely used tools for evaluating financial markets. Therefore, these candlesticks, which show price patterns and differences between buying and selling, provide important data for predicting future price fluctuations. A 15-minute candlestick or 15-minute frame is used. The model showed excellent performance in predicting currency rates (EUR/NZDUSD), with an accuracy based on mean square error (MSE = 1.377e-07). The model also showed better accuracy in predicting sugar prices compared to other models, reaching (MSE = 1.419836). The same results were obtained with the GAS model, where the value was (MSE = 0.000173). This superior performance of the model indicates its ability to generate historical patterns and use them effectively in forecasting financial markets. These results provide promising opportunities for traders and investors to make more guided and intelligent investment decisions based on future trends based on these patterns. By using historical patterns and financial data, LSTM's deep learning model shows exceptional predictive performance. It outperforms traditional machine learning methods such as XGBoost. XGBoost achieved a score on the EUR/NZDUSD exchange rate (MSE = 9.537e-07). The error rate for the presented model is considered to be high. This confirms the success of the represented approach and its ability to enable traders and investors to make more informed and strategic decisions. This ultimately contributes to improving trading conditions and investment outcomes in global financial markets.

Image interpolation is an important topic in the field of image processing. It is defined as the process of transforming low-resolution images into high-resolution ones using image processing methods. Recent studies on interpolation have shown that researchers are focusing on successful interpolation techniques that preserve edge information. Therefore, the edge detection phase plays a vital role in interpolation studies. However, these approaches typically rely on gradient-based linear computations for edge detection. On the other hand, non-linear structures that effectively simulate the human visual system have gained attention. In this study, a non-linear method was developed to detect edge information using a pixel similarity approach. Pixel similarity-based edge detection approach offers both lower computational complexity and more successful interpolation results compared to gradient-based approaches. 1D cubic interpolation was applied to the pixels identified as edges based on pixel similarity, while bicubic interpolation was applied to the remaining pixels. The algorithm was tested on 12 commonly used images and compared with various interpolation techniques. The results were evaluated using metrics such as SSIM and PSNR, as well as visual assessment. The experimental findings clearly demonstrated that the proposed method outperformed other approaches. Additionally, the method offers significant advantages, such as not requiring any parameters and having competitive computational cost.