Differential Privacy with Data Removal for Online Happiness Assessment

Happiness computing aims to assess individuals’ life satisfaction and happiness by analyzing provided data. Currently, online platforms for happiness assessment possess a vast amount of data instead of traditional onsite questionare collection, which carries the risk of privacy breaches. Especially, when a user makes a request to delete some data samples, it is a challenge to balance the utility and efficiency of happiness computing models since it is time-consuming if retraining from scratch without removed data. To this end, we propose Differential Privacy with Data Removal for Online Happiness Assessment, i.e., ADP-CR algorithm. 1) We perform adaptive privacy budget allocation for happiness attributes and propose a layer-wise clipping and dynamic noise addition process for differential privacy training. 2) For trained DP model, we design a certified removal mechanism to handle requests for removing specific data samples from the trained happiness computing model. Experimental results on two widely used datasets demonstrate that, at a certain level of privacy protection, when faced with 1000 data removal requests, the average runtime of our ADP-CR is 13 milliseconds, while retraining takes an average of 118 s, with minimal differences in accuracy between them. Additionally, our ADP-CR shows an average reduction in membership inference attack success rates by 14.25% and 12.75%, compared to other models.