A differential privacy-preserving framework for real-time object detection systems in IoT

With the explosive growth of IoT traffic demand, real-time object detection systems have a large number of computation-intensive tasks, but, limited by their device computing power, the tasks need to be offloaded to edge servers. However, the raw data summarised in the object detection system have much sensitive information, which can cause serious data privacy and security issues once directly released. A novel differential privacy framework DPPF-RODS (Differential Privacy Protection Framework for Real-time Object Detection Systems) is designed for IoT-based real-time object detection systems. The solution addresses critical challenges of sensitive data exposure during computational offloading and inefficient model convergence through a dual adaptive methodology: dynamic noise scaling strategically adjusts privacy protection levels to optimise the balance between data utility and confidentiality, while adaptive training cycles enhance convergence efficiency without compromising detection precision. Three fundamental mechanisms establish comprehensive end-to-end security: constrained local updates maintain consistency across distributed edge devices, noise-injected global aggregation safeguards server-side operations, and optimised public data sampling improves initial model parameters. Finally, a real large-scale dataset is used for evaluation to verify the effectiveness of the DPPF-RODS privacy preserving framework. The result showed that convergence speed of the model has been improved by about 10%.