Eleni Tsalera
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Eleni Tsalera has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Andreas Papadakis
Department of Electrical & Electronic Engineering Educators, School of Pedagogical and Technological Education, 15122 Athens, Greece
9 shared publications - Maria Samarakou
Department of Informatics and Computer Engineering, School of Engineering, University of West Attica, 11521 Athens, Greece
6 shared publications - Ioannis Voyiatzis
Informatics and Computer Engineering, University of West Attica, Athens, Greece, Greece
4 shared publications Maria Samarakou
4 shared publicationsAndreas Papadakis
2 shared publications- Gerasimos Pagiatakis
Department of Electrical & Electronic Engineering Educators, School of Pedagogical and Technological Education, 15122 Athens, Greece
2 shared publications
Latest Publications
- Open Access
- 4 September 2026
- Published by MDPI AG in Future Internet
- Vol. 18 (9) , 469
Evaluation of Vanilla and RAG-Supported LLM Comprehension of the EU AI Act
This paper presents a comparative benchmarking study evaluating the regulatory comprehension of four open-source large language models, TinyLlama-1.1B-Chat, Gemma-2B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, on the EU Artificial Intelligence Act (Regulation EU 2024/1689). A custom benchmark of 100 multiple-choice questions was constructed and classified across three dimensions, namely cognitive difficulty, knowledge type, and regulatory mechanism, with a balanced answer key distribution. Each model was evaluated under vanilla inference, relying exclusively on parametric knowledge, and retrieval-augmented generation (RAG), in which relevant Act passages are retrieved using a FAISS-indexed sentence embedding pipeline with standardized top three chunk retrieval applied uniformly across all models. Vanilla accuracy ranges from 22.0% for TinyLlama-1.1B to 80.0% for Llama-3.1-8B, indicating that larger models perform better. Under the retrieval configuration employed, RAG improves performance of the evaluated models, with gains from 3.0 percentage points for TinyLlama-1.1B to 17.0 for Gemma-2B and 10.0 for both 7B–8B models. Questions related to procedural knowledge, governance and enforcement emerge as weaknesses in the baseline, vanilla setting, partially mitigated by retrieval. Qualitative analysis identifies two RAG failure modes: retrieval failures, where the embedding mechanism returns informationally insufficient passages, and integration failures, where the correct passage is retrieved but not correctly exploited. The study contributes an empirical characterization of small and medium language model regulatory comprehension and a reusable 100-question benchmark.
- 4 September 2026
- Published by MDPI AG in Future Internet
- Vol. 18 (9) , 469
Open AccessEvaluation of Vanilla and RAG-Supported LLM Comprehension of the EU AI Act
This paper presents a comparative benchmarking study evaluating the regulatory comprehension of four open-source large language models, TinyLlama-1.1B-Chat, Gemma-2B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3, on the EU Artificial Intelligence Act (Regulation EU 2024/1689). A custom benchmark of 100 multiple-choice questions was constructed and classified across three dimensions, namely cognitive difficulty, knowledge type, and regulatory mechanism, with a balanced answer key distribution. Each model was evaluated under vanilla inference, relying exclusively on parametric knowledge, and retrieval-augmented generation (RAG), in which relevant Act passages are retrieved using a FAISS-indexed sentence embedding pipeline with standardized top three chunk retrieval applied uniformly across all models. Vanilla accuracy ranges from 22.0% for TinyLlama-1.1B to 80.0% for Llama-3.1-8B, indicating that larger models perform better. Under the retrieval configuration employed, RAG improves performance of the evaluated models, with gains from 3.0 percentage points for TinyLlama-1.1B to 17.0 for Gemma-2B and 10.0 for both 7B–8B models. Questions related to procedural knowledge, governance and enforcement emerge as weaknesses in the baseline, vanilla setting, partially mitigated by retrieval. Qualitative analysis identifies two RAG failure modes: retrieval failures, where the embedding mechanism returns informationally insufficient passages, and integration failures, where the correct passage is retrieved but not correctly exploited. The study contributes an empirical characterization of small and medium language model regulatory comprehension and a reusable 100-question benchmark.
- Open Access
- 19 October 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (20) , 11207
Efficient and Secure GANs: A Survey on Privacy-Preserving and Resource-Aware Models
Generative Adversarial Networks (GANs) generate synthetic content to support applications such as data augmentation, image-to-image translation, and training models where data availability is limited. Nevertheless, their broader deployment is constrained by limitations in data availability, high computational and energy demands, as well as privacy and security concerns. These factors restrict their scalability and integration in real-world applications. This survey provides a systematic review of research aimed at addressing these challenges. Techniques such as few-shot learning, consistency regularization, and advanced data augmentation are examined to address data scarcity. Approaches designed to reduce computational and energy costs, including hardware-based acceleration and model optimization, are also considered. In addition, strategies to improve privacy and security, such as privacy-preserving GAN architectures and defense mechanisms against adversarial attacks, are analyzed. By organizing the literature into these thematic categories, the review highlights available solutions, their trade-offs, and remaining open issues. Our findings underline the growing role of GANs in artificial intelligence, while also emphasizing the importance of efficient, sustainable, and secure designs. This work not only concentrates the current knowledge but also sets the basis for future research.
- 19 October 2025
- Published by MDPI AG in Applied Sciences
- Vol. 15 (20) , 11207
Open AccessEfficient and Secure GANs: A Survey on Privacy-Preserving and Resource-Aware Models
Generative Adversarial Networks (GANs) generate synthetic content to support applications such as data augmentation, image-to-image translation, and training models where data availability is limited. Nevertheless, their broader deployment is constrained by limitations in data availability, high computational and energy demands, as well as privacy and security concerns. These factors restrict their scalability and integration in real-world applications. This survey provides a systematic review of research aimed at addressing these challenges. Techniques such as few-shot learning, consistency regularization, and advanced data augmentation are examined to address data scarcity. Approaches designed to reduce computational and energy costs, including hardware-based acceleration and model optimization, are also considered. In addition, strategies to improve privacy and security, such as privacy-preserving GAN architectures and defense mechanisms against adversarial attacks, are analyzed. By organizing the literature into these thematic categories, the review highlights available solutions, their trade-offs, and remaining open issues. Our findings underline the growing role of GANs in artificial intelligence, while also emphasizing the importance of efficient, sustainable, and secure designs. This work not only concentrates the current knowledge but also sets the basis for future research.
- Open Access
- 9 September 2025
- Published by MDPI AG in Journal of Sensor and Actuator Networks
- Vol. 14 (5) , 91
Impact Evaluation of Sound Dataset Augmentation and Synthetic Generation upon Classification Accuracy
We investigate the impact of dataset augmentation and synthetic generation techniques on the accuracy of supervised audio classification based on state-of-the-art neural networks used as classifiers. Dataset augmentation techniques are applied upon the raw sound and its transformed image format. Specifically, sound augmentation techniques are applied prior to spectral-based transformation and include time stretching, pitch shifting, noise addition, volume controlling, and time shifting. Image augmentation techniques are applied after the transformation of the sound into a scalogram, involving scaling, shearing, rotation, and translation. Synthetic sound generation is based on the AudioGen generative model, triggered through a series of customized prompts. Augmentation and synthetic generation are applied to three sound categories: (a) human sounds, (b) animal sounds, and (c) sounds of things, with each category containing ten sound classes with 20 samples retrieved from the ESC-50 dataset. Sound- and image-orientated neural network classifiers have been used to classify the augmented datasets and their synthetic additions. VGGish and YAMNet (sound classifiers) employ spectrograms, while ResNet50 and DarkNet53 (image classifiers) employ scalograms. The streamlined AI-based process of augmentation and synthetic generation, enhanced classifier fine-tuning and inference allowed for a consistent, multicriteria-comparison of the impact. Classification accuracy has increased for all augmentation and synthetic generation scenarios; however, the increase has not been uniform among the techniques, the sound types, and the percentage of the training set population increase. The average increase in classification accuracy ranged from 2.05% for ResNet50 to 9.05% for VGGish. Our findings reinforce the benefit of audio augmentation and synthetic generation, providing guidelines to avoid accuracy degradation due to overuse and distortion of key audio features.
- 9 September 2025
- Published by MDPI AG in Journal of Sensor and Actuator Networks
- Vol. 14 (5) , 91
Open AccessImpact Evaluation of Sound Dataset Augmentation and Synthetic Generation upon Classification Accuracy
We investigate the impact of dataset augmentation and synthetic generation techniques on the accuracy of supervised audio classification based on state-of-the-art neural networks used as classifiers. Dataset augmentation techniques are applied upon the raw sound and its transformed image format. Specifically, sound augmentation techniques are applied prior to spectral-based transformation and include time stretching, pitch shifting, noise addition, volume controlling, and time shifting. Image augmentation techniques are applied after the transformation of the sound into a scalogram, involving scaling, shearing, rotation, and translation. Synthetic sound generation is based on the AudioGen generative model, triggered through a series of customized prompts. Augmentation and synthetic generation are applied to three sound categories: (a) human sounds, (b) animal sounds, and (c) sounds of things, with each category containing ten sound classes with 20 samples retrieved from the ESC-50 dataset. Sound- and image-orientated neural network classifiers have been used to classify the augmented datasets and their synthetic additions. VGGish and YAMNet (sound classifiers) employ spectrograms, while ResNet50 and DarkNet53 (image classifiers) employ scalograms. The streamlined AI-based process of augmentation and synthetic generation, enhanced classifier fine-tuning and inference allowed for a consistent, multicriteria-comparison of the impact. Classification accuracy has increased for all augmentation and synthetic generation scenarios; however, the increase has not been uniform among the techniques, the sound types, and the percentage of the training set population increase. The average increase in classification accuracy ranged from 2.05% for ResNet50 to 9.05% for VGGish. Our findings reinforce the benefit of audio augmentation and synthetic generation, providing guidelines to avoid accuracy degradation due to overuse and distortion of key audio features.
- Open Access
- 24 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 67-73
Evaluation and Prediction of Resource Usage for multi-parametric Deep Learning training and inference
Deep learning is increasingly used in diverse application fields with results typically surpassing those of traditional machine learning techniques. The portfolio of available neural networks is wide, consisting of the full range in terms of complexity, from compact networks to large ones with multiple layers and parameters. This heterogeneity in the model topology is reflected, not necessarily linearly, on the required computational resources for training and inference. Similarly, the environments where the neural networks are trained and executed are transformed from fully-fledged centralized nodes to distributed architectures with constrained resources. In this view, computational resource requirements can be one of the criteria for resource usage management and neural network selection. In this work we measure the training times for a set of five convolutional neural networks of varying complexity and age (GoogleNet, ShuffleNet, VGGish, YAMNet) under different training configurations, considering the batch size, the number of epochs and the learning rate. These measurements are used to create a CPU-time-training dataset of more than 500 values. This dataset is used to train and evaluate models, based on neural networks, for estimating and predicting training times depending on the models employed and the training parameters. Five regression models have been trained and evaluated in terms of correlation coefficient and root mean square error. In addition, we measure the CPU times needed for inference for a subset of the trained models, which prove to be uncorrelated with the corresponding training times.
- 24 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 67-73
Open AccessEvaluation and Prediction of Resource Usage for multi-parametric Deep Learning training and inference
Deep learning is increasingly used in diverse application fields with results typically surpassing those of traditional machine learning techniques. The portfolio of available neural networks is wide, consisting of the full range in terms of complexity, from compact networks to large ones with multiple layers and parameters. This heterogeneity in the model topology is reflected, not necessarily linearly, on the required computational resources for training and inference. Similarly, the environments where the neural networks are trained and executed are transformed from fully-fledged centralized nodes to distributed architectures with constrained resources. In this view, computational resource requirements can be one of the criteria for resource usage management and neural network selection. In this work we measure the training times for a set of five convolutional neural networks of varying complexity and age (GoogleNet, ShuffleNet, VGGish, YAMNet) under different training configurations, considering the batch size, the number of epochs and the learning rate. These measurements are used to create a CPU-time-training dataset of more than 500 values. This dataset is used to train and evaluate models, based on neural networks, for estimating and predicting training times depending on the models employed and the training parameters. Five regression models have been trained and evaluated in terms of correlation coefficient and root mean square error. In addition, we measure the CPU times needed for inference for a subset of the trained models, which prove to be uncorrelated with the corresponding training times.
- Open Access
- 10 June 2023
- Published by Elsevier in Energy Reports
- Vol. 9, 247-257
CNN-based, contextualized, real-time fire detection in computational resource-constrained environments
- 10 June 2023
- Published by Elsevier in Energy Reports
- Vol. 9, 247-257
Open AccessCNN-based, contextualized, real-time fire detection in computational resource-constrained environments
- Open Access
- 25 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 373-378
CNN-based Segmentation and Classification of Sound Streams under realistic conditions
Audio datasets support the training and validation of Machine Learning algorithms in audio classification problems. Such datasets include different, arbitrarily chosen audio classes. We initially investigate a unifying approach, based on the mapping of audio classes according to the Audioset ontology. Using the ESC-10 audio dataset, a tree-like representation of its classes is created. In addition, we employ an audio similarity calculation tool based on the values of extracted features (spectrum centroid, the spectrum flux and the spectral roll-off). This way the audio classes are connected both semantically and in feature-based manner. Employing the same dataset, ESC-10, we perform sound classification using CNN-based algorithms, after transforming the sound excerpts into images (based on their Mel spectrograms). The YAMNet and VGGish networks are used for audio classification and the accuracy reaches 90%. We extend the classification algorithm with segmentation logic, so that it can be applied into more complex sound excerpts, where multiple sound types are included in a sequential and/or overlapping manner. Quantitative metrics are defined on the behavior of the combined segmentation and segmentation functionality, including two key parameters for the merging operation, the minimum duration of the identified sounds and the intervals. The qualitative metrics are related to the number of sound identification events for a concatenated sound excerpt of the dataset and per each sound class. This way the segmentation logic can operate in a fine- and coarse-grained manner while the dataset and the individual sound classes are characterized in terms of clearness and distinguishability.
- 25 November 2022
- Published by Association for Computing Machinery (ACM)
- p. 373-378
Open AccessCNN-based Segmentation and Classification of Sound Streams under realistic conditions
Audio datasets support the training and validation of Machine Learning algorithms in audio classification problems. Such datasets include different, arbitrarily chosen audio classes. We initially investigate a unifying approach, based on the mapping of audio classes according to the Audioset ontology. Using the ESC-10 audio dataset, a tree-like representation of its classes is created. In addition, we employ an audio similarity calculation tool based on the values of extracted features (spectrum centroid, the spectrum flux and the spectral roll-off). This way the audio classes are connected both semantically and in feature-based manner. Employing the same dataset, ESC-10, we perform sound classification using CNN-based algorithms, after transforming the sound excerpts into images (based on their Mel spectrograms). The YAMNet and VGGish networks are used for audio classification and the accuracy reaches 90%. We extend the classification algorithm with segmentation logic, so that it can be applied into more complex sound excerpts, where multiple sound types are included in a sequential and/or overlapping manner. Quantitative metrics are defined on the behavior of the combined segmentation and segmentation functionality, including two key parameters for the merging operation, the minimum duration of the identified sounds and the intervals. The qualitative metrics are related to the number of sound identification events for a concatenated sound excerpt of the dataset and per each sound class. This way the segmentation logic can operate in a fine- and coarse-grained manner while the dataset and the individual sound classes are characterized in terms of clearness and distinguishability.
- Open Access
- 24 August 2022
- Published by MDPI AG in Applied Sciences
- Vol. 12 (17) , 8455
Feature Extraction with Handcrafted Methods and Convolutional Neural Networks for Facial Emotion Recognition
This research compares the facial expression recognition accuracy achieved using image features extracted (a) manually through handcrafted methods and (b) automatically through convolutional neural networks (CNNs) from different depths, with and without retraining. The Karolinska Directed Emotional Faces, Japanese Female Facial Expression, and Radboud Faces Database databases have been used, which differ in image number and characteristics. Local binary patterns and histogram of oriented gradients have been selected as handcrafted methods and the features extracted are examined in terms of image and cell size. Five CNNs have been used, including three from the residual architecture of increasing depth, Inception_v3, and EfficientNet-B0. The CNN-based features are extracted from the pre-trained networks from the 25%, 50%, 75%, and 100% of their depths and, after their retraining on the new databases. Each method is also evaluated in terms of calculation time. CNN-based feature extraction has proved to be more efficient since the classification results are superior and the computational time is shorter. The best performance is achieved when the features are extracted from shallower layers of pre-trained CNNs (50% or 75% of their depth), achieving high accuracy results with shorter computational time. CNN retraining is, in principle, beneficial in terms of classification accuracy, mainly for the larger databases by an average of 8%, also increasing the computational time by an average of 70%. Its contribution in terms of classification accuracy is minimal when applied in smaller databases. Finally, the effect of two types of noise on the models is examined, with ResNet50 appearing to be the most robust to noise.
- 24 August 2022
- Published by MDPI AG in Applied Sciences
- Vol. 12 (17) , 8455
Open AccessFeature Extraction with Handcrafted Methods and Convolutional Neural Networks for Facial Emotion Recognition
This research compares the facial expression recognition accuracy achieved using image features extracted (a) manually through handcrafted methods and (b) automatically through convolutional neural networks (CNNs) from different depths, with and without retraining. The Karolinska Directed Emotional Faces, Japanese Female Facial Expression, and Radboud Faces Database databases have been used, which differ in image number and characteristics. Local binary patterns and histogram of oriented gradients have been selected as handcrafted methods and the features extracted are examined in terms of image and cell size. Five CNNs have been used, including three from the residual architecture of increasing depth, Inception_v3, and EfficientNet-B0. The CNN-based features are extracted from the pre-trained networks from the 25%, 50%, 75%, and 100% of their depths and, after their retraining on the new databases. Each method is also evaluated in terms of calculation time. CNN-based feature extraction has proved to be more efficient since the classification results are superior and the computational time is shorter. The best performance is achieved when the features are extracted from shallower layers of pre-trained CNNs (50% or 75% of their depth), achieving high accuracy results with shorter computational time. CNN retraining is, in principle, beneficial in terms of classification accuracy, mainly for the larger databases by an average of 8%, also increasing the computational time by an average of 70%. Its contribution in terms of classification accuracy is minimal when applied in smaller databases. Finally, the effect of two types of noise on the models is examined, with ResNet50 appearing to be the most robust to noise.
- Top Cited PapersOpen Access
- 10 December 2021
- Published by MDPI AG in Journal of Sensor and Actuator Networks
- Vol. 10 (4) , 72
Comparison of Pre-Trained CNNs for Audio Classification Using Transfer Learning
The paper investigates retraining options and the performance of pre-trained Convolutional Neural Networks (CNNs) for sound classification. CNNs were initially designed for image classification and recognition, and, at a second phase, they extended towards sound classification. Transfer learning is a promising paradigm, retraining already trained networks upon different datasets. We selected three ‘Image’- and two ‘Sound’-trained CNNs, namely, GoogLeNet, SqueezeNet, ShuffleNet, VGGish, and YAMNet, and applied transfer learning. We explored the influence of key retraining parameters, including the optimizer, the mini-batch size, the learning rate, and the number of epochs, on the classification accuracy and the processing time needed in terms of sound preprocessing for the preparation of the scalograms and spectrograms as well as CNN training. The UrbanSound8K, ESC-10, and Air Compressor open sound datasets were employed. Using a two-fold criterion based on classification accuracy and time needed, we selected the ‘champion’ transfer-learning parameter combinations, discussed the consistency of the classification results, and explored possible benefits from fusing the classification estimations. The Sound CNNs achieved better classification accuracy, reaching an average of 96.4% for UrbanSound8K, 91.25% for ESC-10, and 100% for the Air Compressor dataset.
- 10 December 2021
- Published by MDPI AG in Journal of Sensor and Actuator Networks
- Vol. 10 (4) , 72
Top Cited PapersOpen AccessComparison of Pre-Trained CNNs for Audio Classification Using Transfer Learning
The paper investigates retraining options and the performance of pre-trained Convolutional Neural Networks (CNNs) for sound classification. CNNs were initially designed for image classification and recognition, and, at a second phase, they extended towards sound classification. Transfer learning is a promising paradigm, retraining already trained networks upon different datasets. We selected three ‘Image’- and two ‘Sound’-trained CNNs, namely, GoogLeNet, SqueezeNet, ShuffleNet, VGGish, and YAMNet, and applied transfer learning. We explored the influence of key retraining parameters, including the optimizer, the mini-batch size, the learning rate, and the number of epochs, on the classification accuracy and the processing time needed in terms of sound preprocessing for the preparation of the scalograms and spectrograms as well as CNN training. The UrbanSound8K, ESC-10, and Air Compressor open sound datasets were employed. Using a two-fold criterion based on classification accuracy and time needed, we selected the ‘champion’ transfer-learning parameter combinations, discussed the consistency of the classification results, and explored possible benefits from fusing the classification estimations. The Sound CNNs achieved better classification accuracy, reaching an average of 96.4% for UrbanSound8K, 91.25% for ESC-10, and 100% for the Air Compressor dataset.
- 29 May 2021
- Published by Wiley in Computational Intelligence
- Vol. 37 (4) , 1827-1843
Novel principal component analysis‐based feature selection mechanism for classroom sound classification
Machine learning algorithms for sound classification can be supported by multiple temporal, spectral, and perceptual features extracted from the sound signal. The number of features affects the classification accuracy but also the computational resources requested, so the number of features has to be carefully selected. In this work, we propose a methodology for feature selection based on the principal component analysis. The case study has been the classification of classroom sounds during face‐to‐face module delivery and six sound types have been defined. The proposed method is applied upon a set of 143 sound features to produce feature ranking. The ranking results are compared with those provided by the Relief‐F. Then the selected features are used by five classification algorithms, Linear Discriminant Analysis (LDA), Quadratic Support Vector Machine (QSVM), k Nearest Neighbors, Boosted Trees, and Random Forest. The algorithms are executed with increasing number of features, from 1 to 143, considering both feature rankings, creating 1430 models. The performance of the classification algorithms increases rapidly with the number of features with LDA, QSVM, and Boosted Trees outperforming other methods and surpassing the accuracy ratio of 90% with 25 features.
- 29 May 2021
- Published by Wiley in Computational Intelligence
- Vol. 37 (4) , 1827-1843
Novel principal component analysis‐based feature selection mechanism for classroom sound classification
Machine learning algorithms for sound classification can be supported by multiple temporal, spectral, and perceptual features extracted from the sound signal. The number of features affects the classification accuracy but also the computational resources requested, so the number of features has to be carefully selected. In this work, we propose a methodology for feature selection based on the principal component analysis. The case study has been the classification of classroom sounds during face‐to‐face module delivery and six sound types have been defined. The proposed method is applied upon a set of 143 sound features to produce feature ranking. The ranking results are compared with those provided by the Relief‐F. Then the selected features are used by five classification algorithms, Linear Discriminant Analysis (LDA), Quadratic Support Vector Machine (QSVM), k Nearest Neighbors, Boosted Trees, and Random Forest. The algorithms are executed with increasing number of features, from 1 to 143, considering both feature rankings, creating 1430 models. The performance of the classification algorithms increases rapidly with the number of features with LDA, QSVM, and Boosted Trees outperforming other methods and surpassing the accuracy ratio of 90% with 25 features.
- Open Access
- 1 November 2020
- Published by Elsevier in Energy Reports
- Vol. 6, 223-230
Monitoring, profiling and classification of urban environmental noise using sound characteristics and the KNN algorithm
- 1 November 2020
- Published by Elsevier in Energy Reports
- Vol. 6, 223-230
Open AccessMonitoring, profiling and classification of urban environmental noise using sound characteristics and the KNN algorithm