| In recent years,aerial edge computing(AEC)which combines mobile edge computing(MEC)and UAV clusters has become one of the current research hotspots.AEC combines the advantages of high mobility of UAVs and high real-time performance of MEC,which endows UAV system with the ability to process the collected data in real-time.Applying AEC in post-disaster rescue scenarios,UAVs collect image data from the disaster area and uploads it to the edge server.The edge server uses machine learning algorithms such as object recognition to analyze and process the image data.However,the existing AEC schemes for object recognition in sensitive scenarios have problems such as high communication overhead,data privacy leakage,and weak model robustness.Moreover,they usually require high performance of the UAV edge computing server.The limitations make it difficult to apply AEC for object recognition to the scenario with high-reliability requirements and limited resources,such as post-disaster rescue.To address the above issues,we propose a robust enhanced federated AEC scheme named Fed AEC.Fed AEC assigns the object recognition model training tasks to all UAV nodes,and each UAV node only needs to use the image data collected by itself to perform its model training tasks.Compared with directly transmitting the collected raw image data,the UAV in RFAir only needs to upload the model parameters to the edge server,which greatly reduces the communication overhead between systems and can protect the privacy of the original data to a certain extent.In addition,in view of the problem that low-quality data in the system affect the performance of the model,we designed an asynchronous cyclic storage unlearning mechanism named ACSU adapting to the storage limitation of the UAV.The saved historical model parameters are limited to specified rounds and updated continuously,which greatly saves the storage space of the UAV edge server at the expense of part of the retraining efficiency.ACSU enables the UAV server to find the most recent unaffected round in the storage space after detecting low-quality data,re-aggregate the unaffected local models for that round and continue training.Compared with training from scratch,our mechanism improves the efficiency of retraining.The simulation results show that the calculation amount of Fed AEC in the simulation environment meets the limited computing power of the UAV,and can reduce the amount of communication data by 53.55% to 92.49% compared with the direct transmission of the original image data.And after executing the unlearning mechanism is implemented,the model accuracy can be significantly different from the model accuracy without executing the unlearning mechanism.In addition,recent research works show that pure FL cannot meet the increasing demand for privacy protection.Model parameters may be leaked during the transmission and aggregation process of federated learning,and attackers can further implement inference attacks on the model parameters.Therefore,there is still a risk of privacy leakage in the process of model aggregation.To this end,we design the security model aggregation method SAgg based on secret sharing technology and multi-server settings.SAgg can cut the original model parameters into multiple meaningless shared shares for transmission and aggregation,so that attackers cannot destroy data privacy by stealing model parameters.And SAgg can correctly recover the global model without adversely affecting the accuracy of the model.Compared with other methods,our method only needs to perform local computation without negotiating communication between users,can be deployed simply and quickly,reduces communication overhead between systems.And it’s robust to participant exit operations,so it is more suitable for AEC.Finally,we analyze the security and robustness of SAgg from a theoretical level,then demonstrate the computing performance of SAgg under different numbers of edge servers and the correctness of the recovery model through simulation experiments. |
PDF Full Text Request