Publications

@article{carofilis-vasco_improvement_2023,

title = {Improvement of accent classification models through Grad-Transfer from Spectrograms and Gradient-weighted Class Activation Mapping},

author = {Andrés Carofilis-Vasco and Enrique Alegre and Eduardo Fidalgo and Laura Fernández-Robles},

url = {https://ieeexplore.ieee.org/abstract/document/10190103},

year  = {2023},

date = {2023-01-01},

journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},

abstract = {This article introduces a new method for accent classification using a descriptor called Grad-Transfer, which is extracted using Gradient-weighted Class Activation Mapping (Grad-CAM) based on convolutional neural network (CNN) interpretability. The proposed methodology transfers the knowledge gained by CNNs to classical machine learning algorithms. The study shows that Grad-CAM highlights key regions of spectrograms important for accent prediction, and the generated Grad-Transfer descriptors effectively distinguish different accents. Experiments on the Voice Cloning Toolkit dataset demonstrate an improvement in accent classification accuracy and recall when using Grad-Transfer, outperforming models trained directly on spectrograms.},

note = {Publisher: IEEE},

keywords = {Accent Classification, deep learning, Grad-Transfer, machine learning},

pubstate = {published},

tppubtype = {article}

}

@article{carofilis-vasco_impact_2023,

title = {Impact of facial occlusions in age estimation algorithms for forensic applications},

author = {Andrés Carofilis-Vasco and Deisy Chaves and Alicia Martínez-Mendoza and Eduardo Fidalgo and Víctor González-Castro and Enrique Alegre},

url = {https://buleria.unileon.es/handle/10612/20673},

year  = {2023},

date = {2023-01-01},

journal = {Actas de las VIII Jornadas Nacionales de Investigación en Ciberseguridad: Vigo, 21 a 23 de junio de 2023},

pages = {497–498},

abstract = {This study examines age estimation in Child Sexual Exploitation Materials (CSEM) using deep learning. It evaluates six age estimators on facial images with and without occlusion, focusing on how facial occlusion impacts accuracy. Results show that eye occlusion significantly affects age prediction, especially for minors and older adults. The study offers insights for age estimation in forensic applications related to CSEM.},

note = {Publisher: Universidade de Vigo},

keywords = {Age Estimation, CSEM, deep learning, facial occlusion},

pubstate = {published},

tppubtype = {article}

}

@article{gangwar_triple-biggan_2023,

title = {Triple-BigGAN: Semi-supervised generative adversarial networks for image synthesis and classification on sexual facial expression recognition},

author = {Abhishek Gangwar and Víctor González-Castro and Enrique Alegre and Eduardo Fidalgo},

url = {https://www.sciencedirect.com/science/article/pii/S0925231223000346},

year  = {2023},

date = {2023-01-01},

journal = {Neurocomputing},

volume = {528},

pages = {200–216},

abstract = {This paper introduces a method for automatically recognizing facial expressions that reflect erotic activity, called Sexual Facial Expression Recognition (SFER). The authors present a new dataset, SEA-Faces-30k, consisting of 30,000 manually labeled images categorized as erotic, suggestive-erotic, or non-erotic. Given the difficulty of obtaining large annotated datasets for training deep learning models, they propose a semi-supervised GAN framework, Triple-BigGAN, which learns both a generative model and a classifier simultaneously using unlabeled or partially labeled data. The Triple-BigGAN framework achieves strong performance on the SFER task and performs well on several benchmark datasets. The study also demonstrates that synthetic erotic face images generated by Triple-BigGAN can enhance the performance of deep learning classifiers.},

note = {Publisher: Elsevier},

keywords = {deep learning, Emotion Detection, Facial Expressions, Not Safe For Work, NSFW, Obscene Image Retrieval, Pornography},

pubstate = {published},

tppubtype = {article}

}

@article{joshi_deepsumm_2023,

title = {DeepSumm: Exploiting topic models and sequence to sequence networks for extractive text summarization},

author = {Akanksha Joshi and Eduardo Fidalgo and Enrique Alegre and Laura Fernández-Robles},

url = {https://www.sciencedirect.com/science/article/pii/S0957417422015391},

year  = {2023},

date = {2023-01-01},

journal = {Expert Systems with Applications},

volume = {211},

pages = {118442},

abstract = {This paper introduces DeepSumm, a method for extractive text summarization that combines topic modeling and word embeddings to improve summary quality. DeepSumm uses topic vectors and sequence networks to capture both local and global semantics in a document. It calculates scores for each sentence using Sentence Topic Score (STS), Sentence Content Score (SCS), Sentence Novelty Score (SNS), and Sentence Position Score (SPS), and combines them into a Final Sentence Score (FSS). The method outperforms existing approaches on the DUC 2002 and CNN/DailyMail datasets with improved ROUGE scores.},

note = {Publisher: Pergamon},

keywords = {deep learning, Extractive Summarization, Topic Modeling, Word embedding},

pubstate = {published},

tppubtype = {article}

}

@article{jeuland_assessment_2022,

title = {Assessment of age estimation methods for forensic applications using non-occluded and synthetic occluded facial images},

author = {Elouan Derenee Jeuland and Aitor Del Río Ferreras and Deisy Chaves and Eduardo Fidalgo and Víctor González-Castro and Enrique Alegre},

url = {https://ruc.udc.es/dspace/handle/2183/31412},

year  = {2022},

date = {2022-01-01},

journal = {XLIII Jornadas de Automática},

pages = {972–979},

abstract = {This paper evaluates the performance of six deep-learning-based age estimators for forensic applications, particularly in identifying minors and offenders in Child Sexual Exploitation Materials (CSEM). While deep learning is the state-of-the-art for age estimation, it struggles with minors and older adults due to dataset imbalances. Additionally, offenders often use facial occlusion to obscure identities, further impacting estimator accuracy. The study assesses models on non-occluded and synthetically occluded datasets, revealing that eye occlusion has a greater effect than mouth occlusion. Minors and elderly individuals are the most affected by occlusion, making this research a valuable benchmark for forensic victim profiling.},

note = {Publisher: Universidade da Coruña. Servizo de Publicacións},

keywords = {Age Estimation, CSEM, deep learning, facial occlusion},

pubstate = {published},

tppubtype = {article}

}

@article{bennabhaktula_source_2022,

title = {Source camera device identification from videos},

author = {Guru Swaroop Bennabhaktula and Derrick Timmerman and Enrique Alegre and George Azzopardi},

url = {https://link.springer.com/article/10.1007/s42979-022-01202-0},

year  = {2022},

date = {2022-01-01},

journal = {SN Computer Science},

volume = {3},

number = {4},

pages = {316},

abstract = {This paper addresses the problem of source camera identification in digital videos using deep learning. The authors evaluate different models for camera identification, showing that traditional scene-suppression techniques don't improve performance. They achieved state-of-the-art accuracy on the VISION and QUFVD datasets, outperforming previous methods. The proposed approach does not require flat frames, unlike traditional PRNU-based methods, and is more efficient, making it suitable for use by Law Enforcement Agencies (LEAs).},

note = {Publisher: Springer Nature Singapore Singapore},

keywords = {Camera Recognition, deep learning, Digital Image Forensics, Manchine Learning, Source Camera Identification},

pubstate = {published},

tppubtype = {article}

}

@article{blanco-medina_detecting_2021,

title = {Detecting vulnerabilities in critical infrastructures by classifying exposed industrial control systems using deep learning},

author = {Pablo Blanco-Medina and Eduardo Fidalgo and Enrique Alegre and Andrés Carofilis-Vasco and Francisco Jáñez-Martino and Víctor Fidalgo-Villar},

url = {https://www.mdpi.com/2076-3417/11/1/367},

year  = {2021},

date = {2021-01-01},

urldate = {2021-01-01},

journal = {Applied Sciences},

volume = {11},

number = {1},

pages = {367},

abstract = {This paper presents a deep learning pipeline to classify industrial control panel screenshots into three categories: internet technologies, operation technologies, and others. Using the CRINF-300 dataset, the authors compared CNN architectures and found that Inception-ResNet-V2 and VGG16 performed best, while MobileNet-V1 was recommended for time-sensitive systems with GPU availability.},

note = {Publisher: MDPI},

keywords = {deep learning, Fine-tuning, Image classification, Industrial Control System, Transfer Learning},

pubstate = {published},

tppubtype = {article}

}

@article{fidalgo_state_2021,

title = {State of the Art: Content-based and Hybrid Phishing},

author = {Felipe Castaño and Eduardo Fidalgo and Enrique Alegre and Deisy Chaves and Manuel Sánchez-Paniagua},

url = {https://arxiv.org/abs/2101.12723},

year  = {2021},

date = {2021-01-01},

urldate = {2021-01-01},

abstract = {Phishing attacks have evolved and increased over time and, for this reason, the task of distinguishing between a legitimate site and a phishing site is more and more difficult, fooling even the most expert users. The main proposals focused on addressing this problem can be divided into four approaches: List-based, URL based, content-based, and hybrid. In this state of the art, the most recent techniques using web content-based and hybrid approaches for Phishing Detection are reviewed and compared.},

keywords = {Content-based Features, Cybersecurity, deep learning, Hybrid Features, Hybrid Phishing, machine learning, phishing detection},

pubstate = {published},

tppubtype = {article}

}

@article{saikia_colour_2021,

title = {Colour Neural Descriptors for Instance Retrieval Using CNN Features and Colour Models},

author = {Surajit Saikia and Laura Fernández-Robles and Eduardo Fidalgo and Enrique Alegre},

url = {https://ieeexplore.ieee.org/abstract/document/9344701},

year  = {2021},

date = {2021-01-01},

journal = {IEEE Access},

volume = {9},

pages = {23218–23234},

abstract = {This paper presents color neural descriptors for image retrieval, using CNN features from different color spaces without fine-tuning. An object detector enhances feature extraction, and a stride-based query expansion improves multi-view retrieval. The method achieves state-of-the-art results on multiple datasets.},

note = {Publisher: IEEE},

keywords = {CNN Features, Color Descriptors, deep learning, Image Retrieval, Object Detection},

pubstate = {published},

tppubtype = {article}

}

@article{chaves_assessment_2020,

title = {Assessment and estimation of face detection performance based on deep learning for forensic applications},

author = {Deisy Chaves and Eduardo Fidalgo and Enrique Alegre and Rocío Alaiz-Rodríguez and Francisco Jáñez-Martino and George Azzopardi},

url = {https://www.mdpi.com/1424-8220/20/16/4491},

year  = {2020},

date = {2020-01-01},

journal = {Sensors},

volume = {20},

number = {16},

pages = {4491},

abstract = {This paper explores the effectiveness of deep learning-based face recognition as a forensic tool for criminal investigations. The authors evaluate the speed–accuracy tradeoff of three popular face detectors on the WIDER Face and UFDD datasets using different CPUs and GPUs. They develop a regression model to estimate performance in terms of processing time and accuracy, which could assist forensic laboratories in selecting optimal detection options. Experimental results suggest that the best tradeoff is achieved with 50% image resizing on GPUs and 25% on CPUs, while their regression model achieves a Mean Absolute Error (MAE) of 0.113, demonstrating its potential for forensic applications.},

note = {Publisher: MDPI},

keywords = {deep learning, face recognition, forensic science, speed-accuracy tradeoff},

pubstate = {published},

tppubtype = {article}

}

@article{bennabhaktula_device-based_2020,

title = {Device-based image matching with similarity learning by convolutional neural networks that exploit the underlying camera sensor pattern noise},

author = {Guru Swaroop Bennabhaktula and Enrique Alegre and Dimka Karastoyanova and George Azzopardi},

url = {https://arxiv.org/abs/2004.11443},

year  = {2020},

date = {2020-01-01},

journal = {arXiv preprint arXiv:2004.11443},

abstract = {This paper addresses the challenge of identifying whether two images originate from the same camera, aiding forensic investigations. A two-part network is proposed to quantify the likelihood of a shared source, evaluated on the Dresden dataset (1851 images from 31 cameras). While not yet forensics-ready, the approach achieves 85% accuracy, showing promising results. This research is part of the EU-funded 4NSEEK project focused on combating child sexual abuse.},

keywords = {4NSEEK project, Camera Identification, deep learning, DigitalImage Forensics},

pubstate = {published},

tppubtype = {article}

}

@article{joshi_summcoder_2019,

title = {SummCoder: An unsupervised framework for extractive text summarization based on deep auto-encoders},

author = {Akanksha Joshi and Eduardo Fidalgo and Enrique Alegre and Laura Fernández-Robles},

url = {https://www.sciencedirect.com/science/article/pii/S0957417419302192},

year  = {2019},

date = {2019-01-01},

journal = {Expert Systems with Applications},

volume = {129},

pages = {200–215},

abstract = {This paper introduces SummCoder, a method for extractive text summarization using three metrics: content relevance, novelty, and position relevance. The model performs well on datasets like DUC 2002, Blog Summarization, and a new dataset, TIDSumm, focused on web documents from the Tor network. SummCoder outperforms or matches state-of-the-art methods based on ROUGE metrics, providing useful applications for Law Enforcement Agencies.},

note = {Publisher: Pergamon},

keywords = {Auto-Encoder, deep learning, Extractive Summarization, Extractive Text Summarization, Sentence Embedding, Tor Darknet},

pubstate = {published},

tppubtype = {article}

}

@article{chaves_improving_2019,

title = {Improving Speed-Accuracy Trade-off in Face Detectors for Forensic Tools by Image Resizing},

author = {Deisy Chaves and Eduardo Fidalgo and Enrique Alegre and Pablo Blanco-Medina},

url = {https://www.researchgate.net/profile/Deisy-Chaves/publication/345312886_Improving_Speed-Accuracy_Trade-off_in_FaceDetectors_for_Forensic_Tools_by_Image_Resizing/links/5fa39e3b92851cc28695f323/Improving-Speed-Accuracy-Trade-off-in-FaceDetectors-for-Forensic-Tools-by-Image-Resizing.pdf},

year  = {2019},

date = {2019-01-01},

journal = {V Jornadas Nacionales De Investigación en Ciberseguridad},

volume = {1},

number = {1},

pages = {222–223},

abstract = {This study proposes a strategy to improve face detection speed and accuracy in forensic material analysis, especially for Child Sexual Abuse cases. The method focuses on image resizing to enhance the performance of three deep-learning-based face detectors while ensuring real-time application capability. The results show that resizing images to 50% of their original size strikes the best balance between speed and accuracy, allowing faster face detection with minimal loss in performance.},

note = {Publisher: Universidad de Cáceres},

keywords = {CSA, deep learning, face detection, Forensic Images},

pubstate = {published},

tppubtype = {article}

}

@article{merayo-alba_use_2019,

title = {Use of natural language processing to identify inappropriate content in text},

author = {Sergio Merayo-Alba and Eduardo Fidalgo and Víctor González-Castro and Rocío Alaiz-Rodríguez and Javier Velasco-Mata},

url = {https://link.springer.com/chapter/10.1007/978-3-030-29859-3_22},

year  = {2019},

date = {2019-01-01},

journal = {Hybrid Artificial Intelligent Systems: 14th International Conference, HAIS 2019, León, Spain, September 4–6, 2019, Proceedings 14},

pages = {254–263},

abstract = {The quick development of communication through new technology media such as social networks and mobile phones has improved our lives. However, this also produces collateral problems such as the presence of insults and abusive comments. In this work, we address the problem of detecting violent content on text documents using Natural Language Processing techniques. Following an approach based on Machine Learning techniques, we have trained six models resulting from the combinations of two text encoders, Term Frequency-Inverse Document Frequency and Bag of Words, together with three classifiers: Logistic Regression, Support Vector Machines and Naïve Bayes. We have also assessed StarSpace, a Deep Learning approach proposed by Facebook and configured to use a Hit@1 accuracy. We evaluated these seven alternatives in two publicly available datasets from the Wikipedia Detox Project: Attack and Aggression. StarSpace achieved an accuracy of 0.938 and 0.937 in these datasets, respectively, being the algorithm recommended to detect violent content on text documents among the alternatives evaluated.},

note = {Publisher: Springer International Publishing},

keywords = {deep learning, machine learning, Natural Language Processing, Text Encoders, Violent Content Detection},

pubstate = {published},

tppubtype = {article}

}

@article{chaves_systematic_2018,

title = {A systematic review on object localisation methods in images},

author = {Deisy Chaves and Surajit Saikia and Laura Fernández-Robles and Enrique Alegre and María Trujillo},

url = {https://polipapers.upv.es/index.php/RIAI/article/view/10229},

year  = {2018},

date = {2018-01-01},

journal = {Revista Iberoamericana de Automática e Informática Industrial},

volume = {15},

number = {3},

pages = {231–242},

abstract = {This article provides a systematic review of methods for precise object localization in images, covering techniques from traditional sliding window methods (e.g., Viola-Jones) to modern deep learning-based approaches like Faster-RCNN and Mask-RCNN. It discusses the advantages, disadvantages, and applications of these methods in fields such as industrial inspection, clinical diagnosis, and obstacle detection in vehicles and robots. The review offers an organized summary of these techniques and highlights future research directions.},

note = {Publisher: UNIV POLITECNICA VALENCIA, EDITORIAL UPV CAMINO VERA SN, VALENCIA, 46022, SPAIN},

keywords = {automated detection, Computer vision, deep learning, Faster-RCCN, image processing, Mask-RCNN, object localization, visual inspection},

pubstate = {published},

tppubtype = {article}

}

@article{gangwar_pornography_2017,

title = {Pornography and child sexual abuse detection in image and video: A comparative evaluation},

author = {Abhishek Gangwar and Eduardo Fidalgo and Enrique Alegre and Víctor González-Castro},

url = {https://digital-library.theiet.org/doi/10.1049/ic.2017.0046},

year  = {2017},

date = {2017-01-01},

abstract = {This paper reviews automatic detection methods for pornography and Child Sex Abuse (CSA) material, particularly in sensitive environments like educational or work settings. It evaluates five pornography detection approaches, including traditional skin detection and modern deep learning techniques, using two publicly available pornographic databases. The study finds that methods utilizing multiple features perform better than those relying on single features and that deep learning-based methods outperform traditional approaches, achieving state-of-the-art results. Additionally, the methods were tested on real-world CSA material provided by the Spanish Police.},

note = {Publisher: IET Digital Library},

keywords = {CSA, deep learning, Image classification, pornography detection},

pubstate = {published},

tppubtype = {article}

}

@article{gangwara_triple-biggan_nodate,

title = {Triple-BigGAN: A Semi-Supervised GAN for Image synthesis and classification applied to detect facial sexual expressions},

author = {Abhishek Gangwar and Víctor González-Castro and Enrique Alegre and Eduardo Fidalgo},

url = {https://www.researchgate.net/profile/Eduardo-Fidalgo-4/publication/367056606_Triple-BigGAN_Semi-supervised_Generative_Adversarial_Networks_for_Image_Synthesis_and_Classification_on_Sexual_Facial_Expression_Recognition/links/63c54588d9fb5967c2e00368/Triple-BigGAN-Semi-supervised-Generative-Adversarial-Networks-for-Image-Synthesis-and-Classification-on-Sexual-Facial-Expression-Recognition.pdf},

keywords = {deep learning, Emotion Detection, Facial Expressions, Not Safe For Work, NSFW, Obscene Image Retrieval, Pornography},

pubstate = {published},

tppubtype = {article}

}

@article{carofilis-vasco_classifying_nodate,

title = {Classifying Screenshots of Industrial Control System Using Transfer Learning and Fine-Tuning},

author = {Andrés Carofilis-Vasco and Pablo Blanco-Medina and Francisco Jáñez-Martino and Guru Swaroop Bennabhaktula and Eduardo Fidalgo and Alejandro Prieto-Castro and Víctor Fidalgo-Villar},

url = {https://buleria.unileon.es/handle/10612/20274},

abstract = {This paper proposes a deep learning pipeline to classify industrial control panel screenshots into IT, OT, and other categories. Using transfer learning on nine pre-trained CNNs, the model is tested on the CRINF-300 dataset. Inception-ResNet-V2 achieves the best F1-score (98.32%), while MobileNet-V1 offers the best speed-performance balance.},

keywords = {deep learning, Fine-tuning, Image classification, Industrial Control Systems, Transfer Learning},

pubstate = {published},

tppubtype = {article}

}