Federated Learning and Differential Privacy in AI-Based Surveillance Systems Model
Jason Adiwijaya, Venansius Reynardi Tanaya, Anderies, Andry Chowanda · 2023
Surveillance systems, particularly AI-based ones, play a crucial role in enhancing public security. However, they have also raised significant concerns about privacy due to the sensitive nature of the collected and analyzed data. This research highlights the importance of data privacy without compromising the accuracy of AI-based surveillance systems. We propose the integration of two privacy-preserving methods, Federated Learning (FL) and Differential Privacy (DP), to enhance the security and privacy of data used in the training and testing processes. Federated Learning enables the training of AI models using distributed data without needing to centralize all the data on a single device. In contrast, Differential Privacy offers a statistical framework that safeguards individual data points throughout the training process. Both methods were implemented individually and combined on the chosen AI model. Our findings reveal a trade-off between privacy preservation and model accuracy when applying these privacy-preserving techniques to an AI surveillance system. While there was a reduction in overall accuracy, privacy preservation capabilities yielded satisfactory results. Notably, the optimal results were obtained when combining Federated Learning and Differential Privacy with a noise multiplier of 0.6 and 50 devices participating in the federated learning process, yielded an overall solid accuracy of 69.33, and maintained acceptable confidence levels across different categories. This configuration underscores the need for a careful balance between privacy protection and model performance in AI surveillance systems, opening avenues for future research on refining these techniques. Our code is available at: https://github.com/slimmyYer211/RMCS-PPCV_01