Elena J. Da Costa

Elena J. Da Costa

Auto-generated profile, based on publicly available information

Elena J. Da Costa

Department of Electrical Engineering, University of Braslia, Braslia, Brazil

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
21 January 2026 in IEEE Access
Verified
Journal Article
Open access
JDJaqueline Damacena Duarte
Jaqueline Damacena Duarte
GCGuilherme D. Cândido
Guilherme D. Cândido
JFJosé Ricardo A. De Britto Filho
José Ricardo A. De Britto Filho
JNJoão Souza Neto
João Souza Neto
ECElena J. Da Costa
Elena J. Da Costa
Laerte Peotta MeloLaerte Peotta Melo
Laerte Peotta Melo
Large Language Models (LLMs) are increasingly integrated into various infrastructure and interactive applications. However, their inherent linguistic flexibility introduces security vulnerabilities, particularly through Prompt Injection (PI) attacks. This systematic review provides a comprehensive synthesis of the evolving landscape of PI attacks, including their development trends, classification schemes, evaluation methodologies, defense strategies, and areas for future research. We analyze the evolution of attacks from basic natural language overrides to complex multi-turn manipulations, indirect injections using structured formats such as JSON and XML, and tool-assisted exploits involving automated prompt generation and multi-modal inputs like images. PI attacks are categorized by the level of manipulation, including character, word, sentence, and semantic levels, and by adversarial intent, such as prompt leaking and harmful content generation. The study analyzes these techniques within different LLM architectures and application contexts, including machine translation and Chain of Thought reasoning. It also reviews common evaluation benchmarks, datasets, and metrics. Mitigation strategies examined include automated red teaming frameworks, input validation, content filtering, and alignment through training. It differentiates itself from previous works by conducting a systematic review specifically focused on immediate injection, organizing its attacks, defenses and research gaps. Finally, this review identifies major research challenges and suggests critical directions for improving LLM safety and robustness.

18 August 2025 in IEEE Access
Verified
Journal Article
Open access
JDJaqueline Damacena Duarte
Jaqueline Damacena Duarte
PJPedro Chagas Junior
Pedro Chagas Junior
JCJoão Paulo Carvalho Lustosa Da Costa
João Paulo Carvalho Lustosa Da Costa
ECElena J. da Costa
Elena J. da Costa
Laerte Peotta MeloLaerte Peotta Melo
Laerte Peotta Melo
TSThiago Erivan da Cunha Silva
Thiago Erivan da Cunha Silva
Phishing attacks remain a critical threat in the digital era, exploiting social engineering tactics to compromise user trust and sensitive information, often resulting in financial loss and identity theft. These attacks typically exploit multiple communication channels to impersonate trusted entities, inducing users to click on and access malicious Resource Locations (URLs). Early detection of these URLs can considerably reduce the volume and impact of these attacks. This paper proposes a machine learning-based framework for early detection of phishing URLs, specifically addressing newly registered and parked domains. Using a curated dataset comprising 211,659 URLs obtained from real-time SSL (Secure Sockets Layer) certificate monitoring, popular domain listings, and phishing incident reports, the methods encompass data pre-processing, feature engineering, and model optimization. A Light Gradient Boosting Machine classifier achieved recall of 96.02% and accuracy of 97.28% on a balanced dataset, validated through 10-fold cross-validation. Additionally, feature selection techniques reduced model complexity while maintaining detection effectiveness, facilitating practical deployment. The proposed frameworks provide a versatile tool for phishing prevention and brand protection, demonstrating potential applicability in sectors such as banking and e-commerce. This work contributes to proactive cybersecurity practices addressing evolving threats within digital ecosystems.

10 October 2025 in IEEE Access
Verified
Journal Article
Open access
HRHernany S. Rocha
Hernany S. Rocha
ECElena J. Da Costa
Elena J. Da Costa
JDJaqueline Damacena Duarte
Jaqueline Damacena Duarte
JCJoão Paulo Carvalho Lustosa Da Costa
João Paulo Carvalho Lustosa Da Costa
Modern slavery remains a critical global challenge, affecting millions of individuals through forced labor and human trafficking. As global supply chains become increasingly complex, corporations face growing legal, financial, and reputational risks related to human rights violations. Ensuring transparency and compliance with regulations such as the UK Modern Slavery Act 2015 has become essential, yet progress is hindered by the lack of any available, domain-specific datasets that enable automated monitoring of legal cases. To address this gap, we introduce the Modern Slavery Supply-Chain Cases (MOSSC) dataset, the first resource designed for multi-label text classification (MLTC) of modern slavery litigation. MOSSC contains 1,882 United Kingdom (UK) court judgments from 2015 to 2025, meticulously annotated with 329 fine-grained legal and factual labels relevant to human rights and supply chain compliance. We perform a comprehensive benchmarking study evaluating ten neural architectures, ranging from Bidirectional Gated Recurrent Units (BiGRU) to domain-adapted Bidirectional Encoder Representations from Transformers (BERT) models, including BERT-ECHR pre-trained on European Court of Human Rights cases. Our results demonstrate that BERT-ECHR achieves the highest performance, with a R-Precision at rank 2 (RP@2) of 0.8830 and a normalized Discounted Cumulative Gain at rank 2 (nDCG@2) of 0.8682. In addition, a lightweight BiGRU combined with Law2Vec embeddings offers an efficient alternative for resource-constrained applications. By providing MOSSC along with trained models and open-source code, we enable the development of artificial intelligence (AI)-driven tools for automated monitoring of supply chain risks. This approach empowers companies, auditors, and regulators to detect patterns of exploitation more effectively, assess legal exposure, and enhance compliance with environmental, social, and governance (ESG) requirements. Our work establishes the first robust baselines for AI-driven legal analytics in this domain, paving the way for greater transparency and ethical practices across global supply chains.

27 March 2024 in IEEE Access
Verified
Journal Article
Open access
ECElena J. da Costa
Elena J. da Costa
ASAntonio Santos da Silva
Antonio Santos da Silva
JVJose Alfredo Ruiz Vargas
Jose Alfredo Ruiz Vargas
Joao Paulo Abreu MaranhaoJoao Paulo Abreu Maranhao
Joao Paulo Abreu Maranhao
GSGiovanni Almeida Santos
Giovanni Almeida Santos
JCJoão Paulo Carvalho Lustosa Da Costa
João Paulo Carvalho Lustosa Da Costa
Visual evoked potential (VEP) plays a crucial role in the diagnosis of nerve diseases and epilepsy. By applying luminous stimulation in different frequencies, neuronal electrical voltage changes are measured in the visual cortex area at the rear of the head using magnetoencephalogram (MEG). Such acquired MEG signals are separated into components using blind source separation (BSS). The MEG measurements are analyzed by medical experts and, based on their subjective interpretation, the VEP components are identified. Supervised machine learning (ML) methods can be applied to identify VEP components. However, these methods require labeled data, which must be generated through the subjective interpretation of medical experts. This can be limiting as medical experts traditionally assume a fixed amount of components. This paper proposes a blind framework to estimate the model order of the MEG measurements and to extract the VEP components. In order to estimate the amount of components, the framework exploits a low computational complexity modified Akaike Information Criterion (AIC) and does not require human intervention. In order to overcome the need for labeled data, we propose three approaches to automatically compare components extracted from MEG measurements with and without stimulation. Since each proposed unsupervised identification approach identifies a set of VEP components, we propose their decision fusion using set operations. The proposed framework does not require any human intervention, and it can be used as a complementary tool to support experts in identifying VEP components. The results are presented in terms of average amplitude spectrum and spectral topography. The proposed framework is validated using real MEG measurements.