Pedro Ricardo Morais Inacio
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Pedro Ricardo Morais Inacio has a total of 79 co-authors. Here are the top co-authors by number of shared publications:
- Mário Marques Freire
Universidade da Beira Interior
34 shared publications - Paulo Monteiro
Instituto de Telecomunicações, University of Aveiro
14 shared publications - Manuela Pereira
Universidade da Beira Interior, Covilh, Portugal
12 shared publications - Joana Cabral Costa
Instituto de Telecomunicaes, University of Beira Interior, Covilh, Portugal
8 shared publications - Joao B. F. Sequeiros
Universidade da Beira Interior
6 shared publications - Tiago Roxo
Instituto de Telecomunicaes, University of Beira Interior, Covilh, Portugal
6 shared publications
Latest Publications
- Open Access
- 24 October 2026
- Published by Springer Nature in Cluster Computing
- Vol. 29 (16) , 895
Malware classification in mobile Android platforms: background, concepts and traditional methods
With technological advances, mobile devices are increasingly functional and widely used, with society becoming dependent on them. Mobile devices are used by all age groups and in different contexts, such as educational, commercial, business, and leisure. Mobile device users often hold personal and confidential data that cannot be exposed, and there is a growing concern that this information could be used maliciously. These devices are potentially susceptible to threats similar to those that affect desktop and laptop computers connected to the internet, so protecting users from harm caused by malicious applications becomes important. Therefore, this survey addresses the problem of malware classification in mobile devices. We focus on devices with the Android operating system, which is used in most of these devices, with a market share of slightly more than 70%. This survey provides background information on mobile Android platforms and related malware and presents an overview of methods for Android malware analysis. Then, it explores previously published traditional methods used for malware classification. We pay particular attention to traditional machine learning and other traditional classification methods and their application using static, dynamic, or hybrid malware analysis. Furthermore, this survey also addresses malware prevention methods and evasion studies. Due to the length of this article, we focus here on traditional classification methods, leaving methods based on deep learning or on large language models for another article.
- 24 October 2026
- Published by Springer Nature in Cluster Computing
- Vol. 29 (16) , 895
Open AccessMalware classification in mobile Android platforms: background, concepts and traditional methods
With technological advances, mobile devices are increasingly functional and widely used, with society becoming dependent on them. Mobile devices are used by all age groups and in different contexts, such as educational, commercial, business, and leisure. Mobile device users often hold personal and confidential data that cannot be exposed, and there is a growing concern that this information could be used maliciously. These devices are potentially susceptible to threats similar to those that affect desktop and laptop computers connected to the internet, so protecting users from harm caused by malicious applications becomes important. Therefore, this survey addresses the problem of malware classification in mobile devices. We focus on devices with the Android operating system, which is used in most of these devices, with a market share of slightly more than 70%. This survey provides background information on mobile Android platforms and related malware and presents an overview of methods for Android malware analysis. Then, it explores previously published traditional methods used for malware classification. We pay particular attention to traditional machine learning and other traditional classification methods and their application using static, dynamic, or hybrid malware analysis. Furthermore, this survey also addresses malware prevention methods and evasion studies. Due to the length of this article, we focus here on traditional classification methods, leaving methods based on deep learning or on large language models for another article.
- Open Access
- 2 October 2026
- Published by Elsevier in Image and Vision Computing
- Vol. 176
When zero query meets adversarial attacks: Feature maps as an offensive and defensive approach
- 2 October 2026
- Published by Elsevier in Image and Vision Computing
- Vol. 176
Open AccessWhen zero query meets adversarial attacks: Feature maps as an offensive and defensive approach
- Open Access
- 6 April 2026
- Published by PeerJ in PeerJ Computer Science
- Vol. 12, e3735
LISArD: learning image similarity to defend against gray-box adversarial attacks
State-of-the-art defense mechanisms are typically evaluated in the context of white-box attacks, which is not realistic, as it assumes the attacker can access the gradients of the target network. To protect against this scenario, Adversarial Training (AT) and Adversarial Distillation (AD) include adversarial examples during the training phase, and Adversarial Purification uses a generative model to reconstruct all the images given to the classifier. This article considers an even more realistic evaluation scenario: gray-box attacks, which assume that the attacker knows the architecture and the dataset used to train the target network, but cannot access its gradients. We provide empirical evidence that models are vulnerable to gray-box attacks and propose Learning Image Similarity Adversarial Defense (LISArD), a defense mechanism that does not require training extra models despite slightly increasing training cost, which provides robustness against gray- and white-box attacks without including AT. Our method constructs a cross-correlation matrix by multiplying the embeddings of noisy (augmented with random Gaussian noise) and clean images, then approximates it to a diagonal matrix while simultaneously promoting correct image classification. This process allows the model to recognize noisy image embeddings as similar to those of clean images, thereby improving its generalization. The results show that LISArD can effectively protect against gray-box attacks across different datasets and can be successfully used in multiple architectures. Additionally, without additional training, LISArD resilience carries over to white-box and black-box scenarios, demonstrating its ability to withstand zero-day attacks ( i.e ., previously unseen adversarial samples). Finally, the results show that state-of-the-art AD models underperform significantly when AT is removed and/or when moving to gray-box settings, highlighting their lack of robustness across various conditions and further emphasizing the importance of LISArD. All the source code is available at https://github.com/Joana-Cabral/LISArD .
- 6 April 2026
- Published by PeerJ in PeerJ Computer Science
- Vol. 12, e3735
Open AccessLISArD: learning image similarity to defend against gray-box adversarial attacks
State-of-the-art defense mechanisms are typically evaluated in the context of white-box attacks, which is not realistic, as it assumes the attacker can access the gradients of the target network. To protect against this scenario, Adversarial Training (AT) and Adversarial Distillation (AD) include adversarial examples during the training phase, and Adversarial Purification uses a generative model to reconstruct all the images given to the classifier. This article considers an even more realistic evaluation scenario: gray-box attacks, which assume that the attacker knows the architecture and the dataset used to train the target network, but cannot access its gradients. We provide empirical evidence that models are vulnerable to gray-box attacks and propose Learning Image Similarity Adversarial Defense (LISArD), a defense mechanism that does not require training extra models despite slightly increasing training cost, which provides robustness against gray- and white-box attacks without including AT. Our method constructs a cross-correlation matrix by multiplying the embeddings of noisy (augmented with random Gaussian noise) and clean images, then approximates it to a diagonal matrix while simultaneously promoting correct image classification. This process allows the model to recognize noisy image embeddings as similar to those of clean images, thereby improving its generalization. The results show that LISArD can effectively protect against gray-box attacks across different datasets and can be successfully used in multiple architectures. Additionally, without additional training, LISArD resilience carries over to white-box and black-box scenarios, demonstrating its ability to withstand zero-day attacks ( i.e ., previously unseen adversarial samples). Finally, the results show that state-of-the-art AD models underperform significantly when AT is removed and/or when moving to gray-box settings, highlighting their lack of robustness across various conditions and further emphasizing the importance of LISArD. All the source code is available at https://github.com/Joana-Cabral/LISArD .
- Open Access
- 13 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (10) , 20797-20810
X-IoMT: A Cross-Layer IoMT Dataset for Anomaly Detection With FGSM- and PGD-Based Adversarial Augmentation
With the growing adoption of the Internet of Medical Things (IoMT), concerns around security, reliability, and anomaly detection across heterogeneous system layers are increasing. The essence of contemporary anomaly detection lies in machine learning (ML), training models on datasets for the desired functionality. Nonetheless, currently available datasets are mostly insufficiently supported with multilayer analysis, as they usually target a single architectural layer or a narrow attack type. To close this gap, we offer X-IoMT, a publicly available dataset designed to enable the research of anomaly detection in IoMT environments. In the Internet of Things (IoT) architecture, the X-IoMT data embodies anomalies from the perception, network, and application layers. Besides, it incorporates adversarial perturbations by fast gradient sign method (FGSM) and projected gradient descent (PGD) under both L2 and L $\infty $ norms with extra adversarial augmentation: both on detection performance under common attacks and on resilience against adversarially manipulated samples. Thus, X-IoMT could be an important tool for evaluating and benchmarking reliable anomaly detection models in IoMT scenarios with cross-layer attack coverage and extra adversarial augmentation. This article contributes to the presentation of the dataset, inherently aiming to support the advancement of more reliable and secure IoMT systems.
- 13 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (10) , 20797-20810
Open AccessX-IoMT: A Cross-Layer IoMT Dataset for Anomaly Detection With FGSM- and PGD-Based Adversarial Augmentation
With the growing adoption of the Internet of Medical Things (IoMT), concerns around security, reliability, and anomaly detection across heterogeneous system layers are increasing. The essence of contemporary anomaly detection lies in machine learning (ML), training models on datasets for the desired functionality. Nonetheless, currently available datasets are mostly insufficiently supported with multilayer analysis, as they usually target a single architectural layer or a narrow attack type. To close this gap, we offer X-IoMT, a publicly available dataset designed to enable the research of anomaly detection in IoMT environments. In the Internet of Things (IoT) architecture, the X-IoMT data embodies anomalies from the perception, network, and application layers. Besides, it incorporates adversarial perturbations by fast gradient sign method (FGSM) and projected gradient descent (PGD) under both L2 and L $\infty $ norms with extra adversarial augmentation: both on detection performance under common attacks and on resilience against adversarially manipulated samples. Thus, X-IoMT could be an important tool for evaluating and benchmarking reliable anomaly detection models in IoMT scenarios with cross-layer attack coverage and extra adversarial augmentation. This article contributes to the presentation of the dataset, inherently aiming to support the advancement of more reliable and secure IoMT systems.
- Open Access
- 16 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Computer Society
- Vol. 6, 1673-1685
Classification of Anomalies in Microservices Using an XGBoost-Based Approach With Data Balancing and Hyperparameter Tuning
Microservice architecture has emerged as a leading paradigm for decomposing large monolithic applications into smaller, autonomous services. Although this approach offers many advantages, its complexity, distributed nature, and substantial scale create significant challenges for monitoring and anomaly detection. The vast volume of generated data further exacerbates computational load and detection latency, complicating the identification of anomalies. This study analyses the impact of data balancing and hyperparameter tuning on anomaly classification in microservices and introduces ADMXGB - Anomaly Detection in Microservices using XGBoost, a XGBoost-based framework tailored for anomaly detection in microservices that seamlessly integrates data balancing with hyperparameter tuning. We propose guidelines to determine appropriate threshold values that balance sensitivity with false positives, and show that the framework is model-agnostic, enabling integration with different machine learning algorithms beyond XGBoost. Validation was performed using a four-stage process encompassing preprocessing, training, validation, and testing. ADMXGB demonstrated improvements in both Accuracy and F1-Score, reaching 99.96% in both metrics on the TraceRCA dataset, outperforming the baseline XGBoost method by a margin of 1.46% in Accuracy and 45.62% in F1-Score. Moreover, ADMXGB achieves reductions in execution time (-86.3%) and memory usage (-21.7%), while maintaining an acceptable CPU overhead. These findings highlight the robustness of ADMXGB in delivering high-accuracy classification in a microservice environment.
- 16 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Computer Society
- Vol. 6, 1673-1685
Open AccessClassification of Anomalies in Microservices Using an XGBoost-Based Approach With Data Balancing and Hyperparameter Tuning
Microservice architecture has emerged as a leading paradigm for decomposing large monolithic applications into smaller, autonomous services. Although this approach offers many advantages, its complexity, distributed nature, and substantial scale create significant challenges for monitoring and anomaly detection. The vast volume of generated data further exacerbates computational load and detection latency, complicating the identification of anomalies. This study analyses the impact of data balancing and hyperparameter tuning on anomaly classification in microservices and introduces ADMXGB - Anomaly Detection in Microservices using XGBoost, a XGBoost-based framework tailored for anomaly detection in microservices that seamlessly integrates data balancing with hyperparameter tuning. We propose guidelines to determine appropriate threshold values that balance sensitivity with false positives, and show that the framework is model-agnostic, enabling integration with different machine learning algorithms beyond XGBoost. Validation was performed using a four-stage process encompassing preprocessing, training, validation, and testing. ADMXGB demonstrated improvements in both Accuracy and F1-Score, reaching 99.96% in both metrics on the TraceRCA dataset, outperforming the baseline XGBoost method by a margin of 1.46% in Accuracy and 45.62% in F1-Score. Moreover, ADMXGB achieves reductions in execution time (-86.3%) and memory usage (-21.7%), while maintaining an acceptable CPU overhead. These findings highlight the robustness of ADMXGB in delivering high-accuracy classification in a microservice environment.
- Open Access
- 23 June 2025
- Published by Elsevier in Internet of Things
- Vol. 33
Federated learning for anomaly detection on Internet of Medical Things: A survey
- 23 June 2025
- Published by Elsevier in Internet of Things
- Vol. 33
Open AccessFederated learning for anomaly detection on Internet of Medical Things: A survey
- 14 June 2025
- Published by Springer Nature
- p. 182-200
Framework and Roadmap for Secure Design and Development of Applications in the Cloud and Mobile Ecosystem
- 14 June 2025
- Published by Springer Nature
- p. 182-200
Framework and Roadmap for Secure Design and Development of Applications in the Cloud and Mobile Ecosystem
- 18 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biometrics, Behavior, and Identity Science
- Vol. 7 (3) , 410-421
BIAS: A Body-Based Interpretable Active Speaker Approach
State-of-the-art Active Speaker Detection (ASD) approaches heavily rely on audio and facial features to perform, which is not a sustainable approach in wild scenarios. Although these methods achieve good results in the standard AVA-ActiveSpeaker set, a recent wilder ASD dataset (WASD) showed the limitations of such models and raised the need for new approaches. As such, we propose BIAS, a model that, for the first time, combines audio, face, and body information, to accurately predict active speakers in varying/challenging conditions. Additionally, we design BIAS to provide interpretability by proposing a novel use for Squeeze-and-Excitation blocks, namely in attention heatmaps creation and feature importance assessment. For a full interpretability setup, we annotate an ASD-related actions dataset (ASD-Text) to finetune a ViT-GPT2 for text scene description to complement BIAS interpretability. The results show that BIAS is state-of-the-art in challenging conditions where body-based features are of utmost importance (Columbia, open-settings, and WASD), and yields competitive results in AVA-ActiveSpeaker, where face is more influential than body for ASD. BIAS interpretability also shows the features/aspects more relevant towards ASD prediction in varying settings, making it a strong baseline for further developments in interpretable ASD models, and is available at https://github.com/Tiago-Roxo/BIAS.
- 18 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biometrics, Behavior, and Identity Science
- Vol. 7 (3) , 410-421
BIAS: A Body-Based Interpretable Active Speaker Approach
State-of-the-art Active Speaker Detection (ASD) approaches heavily rely on audio and facial features to perform, which is not a sustainable approach in wild scenarios. Although these methods achieve good results in the standard AVA-ActiveSpeaker set, a recent wilder ASD dataset (WASD) showed the limitations of such models and raised the need for new approaches. As such, we propose BIAS, a model that, for the first time, combines audio, face, and body information, to accurately predict active speakers in varying/challenging conditions. Additionally, we design BIAS to provide interpretability by proposing a novel use for Squeeze-and-Excitation blocks, namely in attention heatmaps creation and feature importance assessment. For a full interpretability setup, we annotate an ASD-related actions dataset (ASD-Text) to finetune a ViT-GPT2 for text scene description to complement BIAS interpretability. The results show that BIAS is state-of-the-art in challenging conditions where body-based features are of utmost importance (Columbia, open-settings, and WASD), and yields competitive results in AVA-ActiveSpeaker, where face is more influential than body for ASD. BIAS interpretability also shows the features/aspects more relevant towards ASD prediction in varying settings, making it a strong baseline for further developments in interpretable ASD models, and is available at https://github.com/Tiago-Roxo/BIAS.
- Open Access
- 30 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Analysis of the Capability and Training of Chat Bots in the Generation of Rules for Firewall or Intrusion Detection Systems
Large Language Models (LLMs) have the potential to aid in closing the knowledge gap in several specific technical areas, such as cybersecurity, by providing a means to translate instructions defined in natural language into specialized system or software specifications (e.g., firewall rules). The work described herein aims at an evaluation of the capability of LLMs to generate rules for firewall and Intrusion Detection Systems (IDSs). A preliminary assessment has shown that widely available chat bots have limited capability to generate correct rules and that caution is needed when using their outputs for the aforementioned objective. This work explores three fine-tuning approaches to address these limitations, each of them with a different objective and achieving distinct success rates. The first approach aimed at testing how well the model was able to use the knowledge obtained from the prompts when the question was structured differently, achieving a success rate of 89%. The second approach aimed at testing how well the model could link the knowledge obtained from two different prompts and reached a success rate of 61%. The final approach aimed at testing if the model could create complex rules by first learning simple rules, achieving a success rate of 79%. It can be concluded that fine-tuning is sufficient to improve chat bots into creating syntactically and technically correct rules for firewalls and IDSs. Results suggest that the development of a specialized model for as many attacks, firewalls and IDSs can indeed be achieved.
- 30 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessAnalysis of the Capability and Training of Chat Bots in the Generation of Rules for Firewall or Intrusion Detection Systems
Large Language Models (LLMs) have the potential to aid in closing the knowledge gap in several specific technical areas, such as cybersecurity, by providing a means to translate instructions defined in natural language into specialized system or software specifications (e.g., firewall rules). The work described herein aims at an evaluation of the capability of LLMs to generate rules for firewall and Intrusion Detection Systems (IDSs). A preliminary assessment has shown that widely available chat bots have limited capability to generate correct rules and that caution is needed when using their outputs for the aforementioned objective. This work explores three fine-tuning approaches to address these limitations, each of them with a different objective and achieving distinct success rates. The first approach aimed at testing how well the model was able to use the knowledge obtained from the prompts when the question was structured differently, achieving a success rate of 89%. The second approach aimed at testing how well the model could link the knowledge obtained from two different prompts and reached a success rate of 61%. The final approach aimed at testing if the model could create complex rules by first learning simple rules, achieving a success rate of 79%. It can be concluded that fine-tuning is sufficient to improve chat bots into creating syntactically and technically correct rules for firewalls and IDSs. Results suggest that the development of a specialized model for as many attacks, firewalls and IDSs can indeed be achieved.
- 11 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biometrics, Behavior, and Identity Science
- Vol. 7 (1) , 61-70
WASD: A Wilder Active Speaker Detection Dataset
Current Active Speaker Detection (ASD) models achieve good results on cooperative settings with reliable face access using only sound and facial features, which is not suited for less constrained conditions. To demonstrate this limitation of current datasets, we propose a Wilder Active Speaker Detection (WASD) dataset, with increased difficulty by targeting the key components of current ASD: audio and face. Grouped into 5 categories, WASD contains incremental challenges for ASD with tactical impairment of audio and face data, and provides a new source for ASD via subject body annotations. To highlight the new challenges of WASD, we divide it into Easy (cooperative settings) and Hard (audio and/or face are specifically degraded) groups, and assess state-of-the-art models performance in WASD and in the most challenging available ASD dataset: AVA-ActiveSpeaker. The results show that: 1) AVA-ActiveSpeaker prepares models for cooperative settings but not wilder ones (surveillance); and 2) current ASD approaches can not reliably perform in wilder settings, even if trained with challenging data. To prove the importance of body for wild ASD, we propose a baseline that complements body with face and audio information that surpass state-of-the-art models in WASD and Columbia. All contributions are available at https://github.com/Tiago-Roxo/WASD .
- 11 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Biometrics, Behavior, and Identity Science
- Vol. 7 (1) , 61-70
WASD: A Wilder Active Speaker Detection Dataset
Current Active Speaker Detection (ASD) models achieve good results on cooperative settings with reliable face access using only sound and facial features, which is not suited for less constrained conditions. To demonstrate this limitation of current datasets, we propose a Wilder Active Speaker Detection (WASD) dataset, with increased difficulty by targeting the key components of current ASD: audio and face. Grouped into 5 categories, WASD contains incremental challenges for ASD with tactical impairment of audio and face data, and provides a new source for ASD via subject body annotations. To highlight the new challenges of WASD, we divide it into Easy (cooperative settings) and Hard (audio and/or face are specifically degraded) groups, and assess state-of-the-art models performance in WASD and in the most challenging available ASD dataset: AVA-ActiveSpeaker. The results show that: 1) AVA-ActiveSpeaker prepares models for cooperative settings but not wilder ones (surveillance); and 2) current ASD approaches can not reliably perform in wilder settings, even if trained with challenging data. To prove the importance of body for wild ASD, we propose a baseline that complements body with face and audio information that surpass state-of-the-art models in WASD and Columbia. All contributions are available at https://github.com/Tiago-Roxo/WASD .