Abstract:

Personalized quality of service （QoS） prediction is crucial for developing high-quality cloud service system.However, the traditional collaborative filtering method, based on centralized training, presents challenges in protecting user privacy.In order to effectively protect user privacy while obtaining highly accurate prediction effect, a distributed user privacy adjustable personalized QoS prediction model for cloud services （DUPPA） was proposed.The model adopted a “server-multi-user” architecture, in which the server coordinated multiple users, handled multiple users’ requests for uploading model gradients and downloading global model, and maintained global model parameters.To further protect user privacy, a user privacy adjustment strategy was proposed to balance privacy and prediction accuracy by adjusting the initialization proportion of local model parameters and gradient upload proportion.In the local model initialization stage, the user calculated the difference matrix between the local model and the global model, and selected the global model parameters corresponding to the larger elements in the difference matrix to initialize the local model parameters.In the gradient upload stage, the user can select some important gradients to upload to the server to meet the privacy protection requirements of different application scenarios.To evaluate the privacy degree of DUPPA, a data reconstruction attack method was proposed for the distributed matrix factorization model gradient sharing scheme.The experimental results show that when DUPPA sets the gradient upload proportion to 0.1 and the local model parameter initialization proportion to 0.5, the predicted MAE and RMSE are reduced by 1.27% and 0.91%, respectively, compared with the traditional centralized matrix factorization model.Besides, when DUPPA sets the gradient upload proportion to 0.1, the privacy degree is 5 times higher than when the gradient upload proportion is 1.And when DUPPA sets the local model parameter initialization proportion to 0.5, the privacy degree is 3.44 times higher than when the local model parameter initialization proportion is 1.

Key words: cloud service, privacy protection, distributed matrix factorization, quality of service prediction