Aplicação de privacidade diferencial na detecção de ataques em internet das coisas

Abstract

The increasing adoption of the Internet of Things expands security challenges, especially in the use of deep learning for attack detection, as these models are susceptible to sensitive data leaks. This work investigates the application of Differential Privacy as a defense against membership inference attacks, which exploit trained models to recover information from the data used in learning. For this purpose, a differentially private neural classifier was implemented using tinyml, allowing its execution on edge devices with limited resources. The experiments showed that while the baseline model achieved 99% accuracy, the application of DP reduced this value to up to 64%, depending on the level of noise added. However, differential privacy significantly decreased the effectiveness of the inference attack, reducing its accuracy from 70% to 52% in the most restrictive scenarios. Furthermore, tinyml optimization enabled execution up to 10 times faster, reducing memory consumption from 8.35 MB to just 0.05 MB. These findings reinforce the feasibility of differential privacy as a strategy to mitigate leakage risks in deep learning models applied to IoT network security, highlighting an important trade-off between privacy and performance.