Luiz Fernando Rust da Costa Carmo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Luiz Fernando Rust da Costa Carmo has a total of 120 co-authors. Here are the top co-authors by number of shared publications:
Luci Pirmez
18 shared publications- Flavia Coimbra Delicato
Universidade Federal Fluminense, Rio de Janeiro, Brazil
14 shared publications - Igor Leão dos Santos
Centro Federal de Educação Tecnológica Celso Suckow da Fonseca, Brazil
7 shared publications - Raphael Carlos Santos Machado
Clavis Information Security, Aloísio Teixeira Street, 278, Building 3, Room 307, Rio de Janeiro 25250-050, RJ, Brazil
7 shared publications - Wilson S. Melo
National Institute of Metrology, Quality, and Technology (Inmetro), Duque de Caxias, RJ, Brazil
7 shared publications - Claudio Miceli de Farias
Federal University of Rio de Janeiro (UFRJ) Rio de Janeiro Rio de Janeiro State Brazil
6 shared publications
Latest Publications
- 20 July 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (29) , 295003
Enhancing table detection and recognition in legacy calibration certificates for metrological digitalization
Legacy calibration certificates (CCs)–commonly stored as scanned PDFs or semi-structured documents–contain critical metrological information, including measurement values, uncertainty statements, units, and contextual metadata, often organized within complex tabular layouts. However, document-specific challenges such as watermarks, double borders, nested tables, and irregular header spacing hinder automated extraction, creating a major obstacle to the large-scale digital transformation (DT) of historical metrological records and the adoption of digital CCs (DCC). This work presents tablecert, a computer vision framework designed to improve table detection (TD) and table structure recognition (TSR) in legacy CCs. Specifically, You Only Look Once (YOLO) was adapted for TD, whereas the Table Transformer (TATR) was adapted for TSR through the incorporation of Frequency Filter-Based (FreqFilter2D), Boundary Refinement Module, lightweight transformer modules, and Low-Rank Adaptation-based fine-tuning strategies. The proposed approach was evaluated using a dataset comprising 100 000 table instances extracted from CC documents under a controlled training and evaluation protocol. The results demonstrated significant improvements, with average gains of 6% in TD (YOLO), 10.5% in structural recognition (TSR), and 5% in valid data conformity. In addition, the framework improved the handling of the five analyzed structural gaps by 40%, reaching a final average coverage of 62%, while achieving final training F1-scores of 0.9996 (YOLO) and 0.9640 (TATR), and reducing the loss values to 0.3781 (YOLO) and 0.2725 (TATR). These findings demonstrate that targeted architectural adaptations can substantially improve the reliability of extracting tabular metrological information from legacy certificates, thereby supporting scalable, traceable, and machine-readable pathways toward DCC-oriented DT. The end-to-end semantic mapping of all certificate components into standardized and interoperable DCC representations remains part of future work.
- 20 July 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (29) , 295003
Enhancing table detection and recognition in legacy calibration certificates for metrological digitalization
Legacy calibration certificates (CCs)–commonly stored as scanned PDFs or semi-structured documents–contain critical metrological information, including measurement values, uncertainty statements, units, and contextual metadata, often organized within complex tabular layouts. However, document-specific challenges such as watermarks, double borders, nested tables, and irregular header spacing hinder automated extraction, creating a major obstacle to the large-scale digital transformation (DT) of historical metrological records and the adoption of digital CCs (DCC). This work presents tablecert, a computer vision framework designed to improve table detection (TD) and table structure recognition (TSR) in legacy CCs. Specifically, You Only Look Once (YOLO) was adapted for TD, whereas the Table Transformer (TATR) was adapted for TSR through the incorporation of Frequency Filter-Based (FreqFilter2D), Boundary Refinement Module, lightweight transformer modules, and Low-Rank Adaptation-based fine-tuning strategies. The proposed approach was evaluated using a dataset comprising 100 000 table instances extracted from CC documents under a controlled training and evaluation protocol. The results demonstrated significant improvements, with average gains of 6% in TD (YOLO), 10.5% in structural recognition (TSR), and 5% in valid data conformity. In addition, the framework improved the handling of the five analyzed structural gaps by 40%, reaching a final average coverage of 62%, while achieving final training F1-scores of 0.9996 (YOLO) and 0.9640 (TATR), and reducing the loss values to 0.3781 (YOLO) and 0.2725 (TATR). These findings demonstrate that targeted architectural adaptations can substantially improve the reliability of extracting tabular metrological information from legacy certificates, thereby supporting scalable, traceable, and machine-readable pathways toward DCC-oriented DT. The end-to-end semantic mapping of all certificate components into standardized and interoperable DCC representations remains part of future work.
- Open Access
- 16 February 2026
- Published by Springer Nature in Research Square
Tablecert: YOLO and TATR Enhanced Models to Boost Table Detection and Recognition in Legacy Documents
The digital transformation of legacy documents remains challenging, as these documents are often unstructured and contain complex table layouts (e.g., watermarks, spaced headers, closely spaced tables, nested structures, and double borders) that degrade the performance of conventional table detection and recognition systems. We propose a modular, plug-and-play adaptation framework for YOLO-based table detection and Table Transformer (TATR)-based structure recognition, combining parameter-efficient LoRA fine-tuning with lightweight architectural modules (e.g., frequency-domain filtering and structural refinements). We evaluate the framework on a dataset of calibration certificates using a controlled training and evaluation protocol with standard detection and structure metrics. The adapted models outperform their respective baselines, mitigating layout-related challenges, and achieve F1-scores of 0.9999 (YOLO) and 0.9640 (TATR), alongside reduced validation loss. The best YOLO adaptation improves robustness in table detection under challenging visual artifacts, whereas the TATR-V6 yields stronger structural recognition. Finally, we show that the proposed FreqFilter2D module is a promising drop-in component for other computer vision architectures.
- 16 February 2026
- Published by Springer Nature in Research Square
Open AccessTablecert: YOLO and TATR Enhanced Models to Boost Table Detection and Recognition in Legacy Documents
The digital transformation of legacy documents remains challenging, as these documents are often unstructured and contain complex table layouts (e.g., watermarks, spaced headers, closely spaced tables, nested structures, and double borders) that degrade the performance of conventional table detection and recognition systems. We propose a modular, plug-and-play adaptation framework for YOLO-based table detection and Table Transformer (TATR)-based structure recognition, combining parameter-efficient LoRA fine-tuning with lightweight architectural modules (e.g., frequency-domain filtering and structural refinements). We evaluate the framework on a dataset of calibration certificates using a controlled training and evaluation protocol with standard detection and structure metrics. The adapted models outperform their respective baselines, mitigating layout-related challenges, and achieve F1-scores of 0.9999 (YOLO) and 0.9640 (TATR), alongside reduced validation loss. The best YOLO adaptation improves robustness in table detection under challenging visual artifacts, whereas the TATR-V6 yields stronger structural recognition. Finally, we show that the proposed FreqFilter2D module is a promising drop-in component for other computer vision architectures.
- Open Access
- 28 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 206431-206446
Hierarchical Multi-Class and Multi-Label Text Classification for Crime Report: A Traditional Machine Learning Approach
Large amounts of digital data are produced daily through society’s use of government and private companies. Digital transformation contributes to the increasing amount of structured and unstructured data stored in digital media. Organizations build centralized data repositories to store and provide information to business areas, supporting Business Intelligence solutions. Some databases store vast amounts of unstructured data, which must be systematized and classified to meet the data owner’s needs. In criminal incident report systems, each recorded incident must be classified as a specific crime, with hundreds or thousands of categories presented to the responsible officer. This work explores a clustering approach to group categories into a hierarchical tree of classes, enabling the use of Machine Learning (ML) models like XGBoost for automated classification of criminal incident reports narratives. As a case study, the Civil Police of of the State of Rio de Janeiro (SEPOL/RJ) has a database with over 6.5 million records, growing daily from Judicial Police Units (JPU) across the state. Each new report requires manual classification. A hierarchical tree of classes was developed to segment the problem, allowing various XGBoost models for automated classification. The proposed hierarchical model with 80 classes achieved an accuracy of 0.463, outperforming the baseline flat model which reached 0.419, along with a 25.48% reduction in training time. The weighted average F1-score obtained by the hierarchical model was 0.48188, while the baseline model reached 0.44061. The improvement was statistically validated through a Wilcoxon signed-rank test, which yielded a p-value of 0.000010.
- 28 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 206431-206446
Open AccessHierarchical Multi-Class and Multi-Label Text Classification for Crime Report: A Traditional Machine Learning Approach
Large amounts of digital data are produced daily through society’s use of government and private companies. Digital transformation contributes to the increasing amount of structured and unstructured data stored in digital media. Organizations build centralized data repositories to store and provide information to business areas, supporting Business Intelligence solutions. Some databases store vast amounts of unstructured data, which must be systematized and classified to meet the data owner’s needs. In criminal incident report systems, each recorded incident must be classified as a specific crime, with hundreds or thousands of categories presented to the responsible officer. This work explores a clustering approach to group categories into a hierarchical tree of classes, enabling the use of Machine Learning (ML) models like XGBoost for automated classification of criminal incident reports narratives. As a case study, the Civil Police of of the State of Rio de Janeiro (SEPOL/RJ) has a database with over 6.5 million records, growing daily from Judicial Police Units (JPU) across the state. Each new report requires manual classification. A hierarchical tree of classes was developed to segment the problem, allowing various XGBoost models for automated classification. The proposed hierarchical model with 80 classes achieved an accuracy of 0.463, outperforming the baseline flat model which reached 0.419, along with a 25.48% reduction in training time. The weighted average F1-score obtained by the hierarchical model was 0.48188, while the baseline model reached 0.44061. The improvement was statistically validated through a Wilcoxon signed-rank test, which yielded a p-value of 0.000010.
- 20 September 2024
- Published by IOS Press
- p. 332-351
Intelligent Attacks on Cyber-Physical Systems and Critical Infrastructures
This chapter provides an overview of the evolving landscape of attacks in cyber-physical systems (CPS) and critical infrastructures, highlighting the possible use of Artificial Intelligence (AI) algorithms to develop intelligent cyberattacks. It describes various existing methods used to carry out intelligent attacks in Operational Technology (OT) environments and discusses AI-driven tools that automate penetration tests in Information Technology (IT) systems, which could potentially be used as attack tools. The chapter also discusses mitigation strategies to counter these emerging intelligent attacks by hindering the learning process of AI-based attacks and points to future research directions on the matter.
- 20 September 2024
- Published by IOS Press
- p. 332-351
Intelligent Attacks on Cyber-Physical Systems and Critical Infrastructures
This chapter provides an overview of the evolving landscape of attacks in cyber-physical systems (CPS) and critical infrastructures, highlighting the possible use of Artificial Intelligence (AI) algorithms to develop intelligent cyberattacks. It describes various existing methods used to carry out intelligent attacks in Operational Technology (OT) environments and discusses AI-driven tools that automate penetration tests in Information Technology (IT) systems, which could potentially be used as attack tools. The chapter also discusses mitigation strategies to counter these emerging intelligent attacks by hindering the learning process of AI-based attacks and points to future research directions on the matter.
- 8 January 2021
- Published by Springer Nature in Computing
- Vol. 103 (6) , 1105-1120
A multisensor prediction-based heuristic for the internet of things
- 8 January 2021
- Published by Springer Nature in Computing
- Vol. 103 (6) , 1105-1120
A multisensor prediction-based heuristic for the internet of things
- 2 December 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 70 (00189456) , 1-11
Securing Smart Meters Through Physical Properties of Their Components
This paper presents a security strategy to protect smart meters against attacks that compromise critical electronic components. Many of these attacks consist of stealing energy by manipulating smart meters’ measurements. For example, a malicious entity can replace or tamper with components that execute measuring procedures and store sensitive parameters. This type of attack is common on devices like smart meters that are part of environments considered hostile, where attackers have easy access to the device. Therefore, it is necessary to develop techniques in which physical manipulations on smart meters by attackers are unsuccessful. Our strategy uses physical properties from these components to create secure identities for the meter. We present two main contributions. The first one is inspired by Physical Unclonable Functions and extracts unique identifiers from SRAM components. Then we combine these identifiers to create a strong identity. The other contribution uses physical context information from the voltage levels in the smart meter’s power supply to yield dynamic context identities. We also validate our proposals in experiments using a hardware prototype that embeds Arduino® microprocessors, SRAM memories, and voltage sensors. The results show that our idea is suitable for implementation in real smart meters and can help protect these devices from attacks against their components.
- 2 December 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 70 (00189456) , 1-11
Securing Smart Meters Through Physical Properties of Their Components
This paper presents a security strategy to protect smart meters against attacks that compromise critical electronic components. Many of these attacks consist of stealing energy by manipulating smart meters’ measurements. For example, a malicious entity can replace or tamper with components that execute measuring procedures and store sensitive parameters. This type of attack is common on devices like smart meters that are part of environments considered hostile, where attackers have easy access to the device. Therefore, it is necessary to develop techniques in which physical manipulations on smart meters by attackers are unsuccessful. Our strategy uses physical properties from these components to create secure identities for the meter. We present two main contributions. The first one is inspired by Physical Unclonable Functions and extracts unique identifiers from SRAM components. Then we combine these identifiers to create a strong identity. The other contribution uses physical context information from the voltage levels in the smart meter’s power supply to yield dynamic context identities. We also validate our proposals in experiments using a hardware prototype that embeds Arduino® microprocessors, SRAM memories, and voltage sensors. The results show that our idea is suitable for implementation in real smart meters and can help protect these devices from attacks against their components.
- 17 June 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 68 (8) , 7672-7683
Model-Based Stealth Attack to Networked Control System Based on Real-Time Ethernet
Industrial control systems (ICS) include networked control systems (NCS), which use Real-Time Ethernet (RTE) protocols since many years, well before the Time Sensitive Networking (TSN) initiative debut. Today, Ethernet based control systems are used all across Industry 4.0, including in critical applications, allowing for straight integration with IT layers. Even if it is known that current RTE protocols do not have strong authentication or ciphering options, it is still very challenging to perform undetected cyber-attacks to these protocols while the NSC is in operation, in particular because such attacks must comply with very strict and small temporal constraints. In this paper, a model based attack is proposed for service degradation of NCS. The attack is carried out in real-time and it can remain undetected for the entire plant life. The attack can be applied to any RTE protocols and, without loss of generality, a detailed analysis of stealth techniques is provided for a specific real use case based on PROFINET. The experimental results demonstrate the feasibility of the proposed attack and its high effectiveness. The paper also points out some possible future investigation directions in order to mitigate the attack.
- 17 June 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 68 (8) , 7672-7683
Model-Based Stealth Attack to Networked Control System Based on Real-Time Ethernet
Industrial control systems (ICS) include networked control systems (NCS), which use Real-Time Ethernet (RTE) protocols since many years, well before the Time Sensitive Networking (TSN) initiative debut. Today, Ethernet based control systems are used all across Industry 4.0, including in critical applications, allowing for straight integration with IT layers. Even if it is known that current RTE protocols do not have strong authentication or ciphering options, it is still very challenging to perform undetected cyber-attacks to these protocols while the NSC is in operation, in particular because such attacks must comply with very strict and small temporal constraints. In this paper, a model based attack is proposed for service degradation of NCS. The attack is carried out in real-time and it can remain undetected for the entire plant life. The attack can be applied to any RTE protocols and, without loss of generality, a detailed analysis of stealth techniques is provided for a specific real use case based on PROFINET. The experimental results demonstrate the feasibility of the proposed attack and its high effectiveness. The paper also points out some possible future investigation directions in order to mitigate the attack.
- Open Access
Identification of Data Injection Attacks in Networked Control Systems Using Noise Impulse Integration
The benefits of using Networked Control Systems (NCS) in the growing Industry 4.0 are numerous, including better management and operational capabilities, as well as costs reduction. However, despite these benefits, the use of NCSs can also expose physical plants to new threats originated in the cyber domain—such as data injection attacks in NCS links through which sensors and controllers transmit signals. In this sense, this work proposes a link monitoring strategy to identify linear time-invariant (LTI) functions executed during controlled data injection attacks by a Man-in-the-Middle hosted in an NCS link. The countermeasure is based on a bioinspired metaheuristic, called Backtracking Search Optimization Algorithm (BSA), and uses white Gaussian noise to excite the attack function. To increase the accuracy of this countermeasure, it is proposed the Noise Impulse Integration (NII) technique, which is developed using the radar pulse integration technique as inspiration. The results demonstrate that the proposed countermeasure is able to accurately identify LTI attack functions, here executed to impair measurements transmitted by the plant sensor, without interfering with the NCS behavior when the system is in its normal operation. Moreover, the results indicate that the NII technique can increase the accuracy of the attack identification.
Open AccessIdentification of Data Injection Attacks in Networked Control Systems Using Noise Impulse Integration
The benefits of using Networked Control Systems (NCS) in the growing Industry 4.0 are numerous, including better management and operational capabilities, as well as costs reduction. However, despite these benefits, the use of NCSs can also expose physical plants to new threats originated in the cyber domain—such as data injection attacks in NCS links through which sensors and controllers transmit signals. In this sense, this work proposes a link monitoring strategy to identify linear time-invariant (LTI) functions executed during controlled data injection attacks by a Man-in-the-Middle hosted in an NCS link. The countermeasure is based on a bioinspired metaheuristic, called Backtracking Search Optimization Algorithm (BSA), and uses white Gaussian noise to excite the attack function. To increase the accuracy of this countermeasure, it is proposed the Noise Impulse Integration (NII) technique, which is developed using the radar pulse integration technique as inspiration. The results demonstrate that the proposed countermeasure is able to accurately identify LTI attack functions, here executed to impair measurements transmitted by the plant sensor, without interfering with the NCS behavior when the system is in its normal operation. Moreover, the results indicate that the NII technique can increase the accuracy of the attack identification.
- 1 July 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Sensor Data Prediction techniques for nodes in IoT (poster)
As many Internet of Things devices rely on batteries, reducing energy consumption where battery replacement may be a problem may extend their lifetime and possible applications. This problem may be hindering the expansion of IoT devices usage in some applications. In this paper, we propose three data prediction and aggregation techniques to be applied to sensor readings and avoid unnecessary communication between a sensor node and Sink, so as to save energy and reduce network traffic. Since we have limited memory and processing capacity within sensors, this proposed approach collects data, compare the readings with expected value from one of the three techniques, and if the monitored data is within an ex-ante defined and acceptable range, no data transmission is necessary - the Sink would infer the actual reading within a uncertainty margin. Otherwise, last reading and estimated trend is transmitted and both sensor node and Sink are synchronized. The proposed techniques are (i) constant; (ii) linear; and (iii) using machine learning (Weightless Neural Network). We discuss pros and cons of each technique depending on the characteristics of the monitored data, taking into consideration the number of times the sensor reading had to be sent to Sink, and the average difference between estimated and actual reading.
- 1 July 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Sensor Data Prediction techniques for nodes in IoT (poster)
As many Internet of Things devices rely on batteries, reducing energy consumption where battery replacement may be a problem may extend their lifetime and possible applications. This problem may be hindering the expansion of IoT devices usage in some applications. In this paper, we propose three data prediction and aggregation techniques to be applied to sensor readings and avoid unnecessary communication between a sensor node and Sink, so as to save energy and reduce network traffic. Since we have limited memory and processing capacity within sensors, this proposed approach collects data, compare the readings with expected value from one of the three techniques, and if the monitored data is within an ex-ante defined and acceptable range, no data transmission is necessary - the Sink would infer the actual reading within a uncertainty margin. Otherwise, last reading and estimated trend is transmitted and both sensor node and Sink are synchronized. The proposed techniques are (i) constant; (ii) linear; and (iii) using machine learning (Weightless Neural Network). We discuss pros and cons of each technique depending on the characteristics of the monitored data, taking into consideration the number of times the sensor reading had to be sent to Sink, and the average difference between estimated and actual reading.
- Open Access
- 25 February 2018
- Published by Wiley in Security and Communication Networks
- Vol. 2018, 1-14
Using Physical Context-Based Authentication against External Attacks: Models and Protocols
Modern systems are increasingly dependent on the integration of physical processes and information technologies. This trend is remarkable in applications involving sensor networks, cyberphysical systems, and Internet of Things. Despite its complexity, such integration results in physical context information that can be used to improve security, especially authentication. In this paper, we show that entities sharing the same physical context can use it for establishing a secure communication channel and protecting each other against external attacks. We present such approach proposing a theoretical model for generating unique bitstreams. Two different protocols are suggested. Each one is evaluated using probabilistic analysis and simulation. In the end, we implement the authentication mechanism in a case study using networks radio signal as physical event generator. The results demonstrate the performance of each of the protocols and their suitability for applications in real world.
- 25 February 2018
- Published by Wiley in Security and Communication Networks
- Vol. 2018, 1-14
Open AccessUsing Physical Context-Based Authentication against External Attacks: Models and Protocols
Modern systems are increasingly dependent on the integration of physical processes and information technologies. This trend is remarkable in applications involving sensor networks, cyberphysical systems, and Internet of Things. Despite its complexity, such integration results in physical context information that can be used to improve security, especially authentication. In this paper, we show that entities sharing the same physical context can use it for establishing a secure communication channel and protecting each other against external attacks. We present such approach proposing a theoretical model for generating unique bitstreams. Two different protocols are suggested. Each one is evaluated using probabilistic analysis and simulation. In the end, we implement the authentication mechanism in a case study using networks radio signal as physical event generator. The results demonstrate the performance of each of the protocols and their suitability for applications in real world.