Diabetic retinopathy (DR) is a leading cause of vision impairment worldwide, requiring early detection for effective treatment. Deep learning models have been widely used for automated DR classification, with Convolutional Neural Networks (CNNs) being the most established approach. Recently, Vision Transformers (ViTs) have shown promise, but a direct comparison of their performance and interpretability remains limited. Additionally, hybrid models that combine CNN and transformer-based architectures have not been extensively studied. This work systematically evaluates CNNs (ResNet-50), ViTs (Vision Transformer and SwinV2-Tiny), and hybrid models (Convolutional Vision Transformer, LeViT-256, and CvT-13) on DR classification using publicly available retinal image datasets. The models are assessed based on classification accuracy and interpretability, applying Grad-CAM and Attention-Rollout to analyze decision-making patterns. Results indicate that hybrid models outperform both standalone CNNs and ViTs, achieving a better balance between local feature extraction and global context awareness. The best-performing model (CvT-13) achieved a Quadratic Weighted Kappa (QWK) score of 0.84 and an AUC of 0.93 on the test set. Interpretability analysis shows that CNNs focus on fine-grained lesion details, while ViTs exhibit broader but less localized attention. These findings provide valuable insights for optimizing deep learning models in medical imaging, supporting the development of clinically viable AI-driven DR screening systems.

Privacy policies are the main method for informing Internet users of how their data are collected and shared. This study aims to analyze the deficiencies of privacy policies in terms of readability, vague statements, and the use of pacifying phrases concerning privacy. This represents the undertaking of a step forward in the literature on this topic through a comprehensive analysis encompassing both time and website coverage. It characterizes trends across website categories, top-level domains, and popularity ranks. Furthermore, studying the development in the context of the General Data Protection Regulation (GDPR) offers insights into the impact of regulations on policy comprehensibility. The findings reveal a concerning trend: privacy policies have grown longer and more ambiguous, making it challenging for users to comprehend them. Notably, there is an increased proportion of vague statements, while clear statements have seen a decrease. Despite this, the study highlights a steady rise in the inclusion of reassuring statements aimed at alleviating readers’ privacy concerns.

Empirical insights into promising commercial sentiment analysis solutions that go beyond the claims of their vendors are rare. Moreover, due to the constant evolution in the field, previous studies are far from reflecting the current situation. The goal of this article is to evaluate and compare current solutions using two experimental studies. In the first part of the study, based on tweets about airline service quality, we test the solutions of six vendors with different market power, such as Amazon, Google, IBM, Microsoft, Lexalytics, and MeaningCloud, and report their measures of accuracy, precision, recall, (macro)F1, time performance, and service level agreements (SLA). Furthermore, we compare two of the services in depth with multiple data sets and over time. The services tested here are Google Cloud Natural Language API and MeaningCloud Sentiment Analysis API. For evaluating the results over time, we use the same data set as in November 2020. In addition, further topic-specific and general Twitter data sets are used. The experiments show that the IBM Watson NLU and Google Cloud Natural Language API solutions may be preferred when negative text detection is the primary concern. When tested in July 2022, the Google Cloud Natural Language API was still the clear winner compared to the MeaningCloud Sentiment Analysis API, but only on the airline service quality data set; on the other data sets, both services provided specific benefits and drawbacks. Furthermore, we detected changes in the sentiment classification over time with both services. Our results motivate that an independent, critical, and longitudinal experimental analysis of sentiment analysis services can provide interesting insights into their overall reliability and particular classification accuracy beyond marketing claims to critically compare solutions based on real data and analyze potential weaknesses and margins of error before making an investment.

Vibrato is one of the most frequently investigated parameters in singing. Most studies examine a very limited number of sound samples to answer specific questions. In the novel approach of our interdisciplinary vibrato project, a sophisticated database of 1723 sound samples was compiled, containing examples from 30 operas, one oratorio and 73 vocal exercises. Single tones undisturbed by orchestral sound were selected in the vocal registers soprano, tenor, baritone and bass according to a standardized classification system dividing voices into either a lyric or dramatic voice type. Various algorithms were implemented in a Matlab environment to automatically evaluate sound samples, with respect to vibrato extent and rate. An AM/FM demodulation scheme is presented with postprocessing required for automatic calculation of these two parameters. The vibrato extent was higher for the dramatic compared to the lyric voice type in soprano and tenor voices. In contrast, the extent of lyric baritones was significantly higher than for dramatic ones. Moreover, the vibrato rate was slightly lower for the dramatic voice type in all vocal registers. It turned out that the classification of bass voices into lyric and dramatic, unlike the other vocal registers, was not scientifically reproducible with our methods. Further vibrato results regarding aging, intra-individual and general stylistic development over time are presented and their essential relevance for vocal pedagogy are discussed.

Fog computing (FC) is a distributed architecture in which computing resources and services are placed on edge devices closer to data sources. This enables more efficient data processing, shorter latency times, and better performance. Fog computing was shown to be a promising solution for addressing the new computing requirements. However, there are still many challenges to overcome to utilize this new computing paradigm, in particular, reliability and security. Following this need, a systematic literature review was conducted to create a list of requirements. As a result, the following four key requirements were formulated: (1) low latency and response times; (2) scalability and resource management; (3) fault tolerance and redundancy; and (4) privacy and security. Low delay and response can be achieved through edge caching, edge real-time analyses and decision making, and mobile edge computing. Scalability and resource management can be enabled by edge federation, virtualization and containerization, and edge resource discovery and orchestration. Fault tolerance and redundancy can be enabled by backup and recovery mechanisms, data replication strategies, and disaster recovery plans, with a residual number system (RNS) being a promising solution. Data security and data privacy are manifested in strong authentication and authorization mechanisms, access control and authorization management, with fully homomorphic encryption (FHE) and the secret sharing system (SSS) being of particular interest.

Homomorphic encryption (HE) is a promising solution for handling sensitive data in semi-trusted third-party computing environments, as it enables processing of encrypted data. However, applying sophisticated techniques such as machine learning, statistics, and image processing to encrypted data remains a challenge. The computational complexity of some encrypted operations can significantly increase processing time. In this paper, we focus on the analysis of two state-of-the-art HE matrix multiplication algorithms with the best time and space complexities. We show how their performance depends on the libraries and the execution context, considering the standard Cheon–Kim–Kim–Song (CKKS) HE scheme with fixed-point numbers based on the Microsoft SEAL and PALISADE libraries. We show that Windows OS for the SEAL library and Linux OS for the PALISADE library are the best options. In general, PALISADE-Linux outperforms PALISADE-Windows, SEAL-Linux, and SEAL-Windows by 1.28, 1.59, and 1.67 times on average for different matrix sizes, respectively. We derive high-precision extrapolation formulas to estimate the processing time of HE multiplication of larger matrices.