13August2026
09:00 Doctoral defense Room 85 of IC2
Topic on
Phishing Detection in Conversational Platforms
Student
Stephane de Freitas Schwarz
Advisor / Teacher
Anderson de Rezende Rocha
Brief summary
Application-to-Person (A2P) messaging has transformed how companies interact with their customers, enabling scalable, efficient, and personalized communication across various sectors. However, the widespread adoption, low cost, and accessibility of messaging services have also contributed to an increase in fraudulent activities, particularly smishing, one of the oldest and most prevalent forms of phishing targeting short message service (SMS). This challenge has been intensified with the emergence of generative language models, which have expanded both the scale and sophistication of fraudulent message generation, allowing attackers to produce highly persuasive content while evading traditional detection mechanisms. Fraudulent messages impact multiple institutions throughout the ecosystem, including messaging platforms responsible for content integrity, legitimate businesses whose brands can be imitated and harmed, and end users, who remain the ultimate targets of these attacks. Although previous approaches have focused primarily on the continuous adaptation of supervised detection models on devices through large-scale fine-tuning, little attention has been paid to exploiting campaign-level structures, despite the significant overlap between messages distributed across different campaigns. Furthermore, deployments in real-world scenarios impose requirements that go beyond detection accuracy, especially low-latency response times that avoid introducing delays in the message delivery pipeline. This thesis addresses these challenges by bridging the gap between the robustness of models and industry requirements for detecting malicious content in SMS environments, with extensions applicable to other messaging channels, including over-the-top (OTT) and Rich Communication Services (RCS). We investigated methods ranging from simple heuristic approaches to more complex machine learning techniques, analyzing their advantages and limitations to develop an integrated detection framework that exploits multiple features throughout the entire delivery pipeline. To address the scarcity of high-quality datasets in the literature, we collaborated with the Sinch communication platform to build a dataset and explored a zero-shot approach to accelerate the labeling process. Our results show that an integrated solution, combining multiple sources of contextual information, can better adapt, in real time, to emerging fraud patterns. Although the adversarial nature of fraud detection remains a continuous game of cat and mouse, this work demonstrates that the combination of robust modeling and contextual intelligence can significantly strengthen defenses against malicious messages, while also highlighting opportunities for future improvements as adversaries continue to evolve.
Examination Board
Headlines:
| Anderson de Rezende Rocha | IC / UNICAMP |
| Thiago Alexandre Salgueiro Pardo | ICMC / USP |
| Hugo Pedro Martins Carriço Proença | UBI/Portugal |
| Washington Luiz Miranda da Cunha | IC / UNICAMP |
| Helena de Almeida Maia | IC / UNICAMP |
Substitutes:
| João Paulo Papa | FC / UNESP |
| David Menotti Gomes | DInf / UFPR |
| Marjory Cristiany da Costa Abreu | SHU/United Kingdom |