Abordagem integrada de deep learning e p4 para a detecção e mitigação de ataques na borda de redes IoT

• Ano de defesa: 2025
• Autor(a) principal: Silva, Antônia Mayara de Almeida da
• Orientador(a): Bonfim, Michel Sales
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Área do conhecimento CNPq: CNPQ: CIÊNCIAS EXATAS E DA TERRA: CIÊNCIA DA COMPUTAÇÃO
• Link de acesso: http://repositorio.ufc.br/handle/riufc/80317

The Internet of Things is a fundamental topic when talking about technological advances in last years. Its wide use arises from the immense variety of applications it encompasses, such as smart cities and its use in industry. However, amidst so many aspects, it is essential to dedicate special attention to the security area. As an alternative, Artificial Intelligence models are often implemented for detecting and preventing attacks. In this work, we present a solution that utilizes a deep learning algorithm as a technique to detect attacks on IoT networks. Additionally, we address the advantage of mitigating detected attacks quickly and efficiently through the implementation of P4, an innovative language that enables data plane programmability. We also discuss attacks targeting learning models directly, confusing their inferences and compromising their efficiency, known as adversarial attacks. As initial results, we trained and tested three DL algorithms: CNN, GRU, and LSTM, using the Edge-IIoTset dataset. The results showed an accuracy value above 99% for all three algorithms. At this stage, precision, recall, and F1 score metrics were also evaluated. We further implemented the three algorithms on a Raspberry Pi 4 equipped with the P4Pi platform. For this implementation, we used the OpenVINO toolkit to convert and execute the algorithms. In subsequent steps, we evaluated the algorithms’ performance on the Raspberry Pi, analyzing network latency, CPU and memory consumption, and packet processing time. We introduced four types of adversarial attacks (FGSM, PGD, CW, and GAN) into the algorithms and assessed their impact. Finally, we retrained the models by including these attacks in the test set so that the algorithms could detect them accurately, and we selected, based on the overall analysis of the metrics, the CNN algorithm for the final implementation in the solution.