Double Insurance: Incentivized Federated Learning with Differential Privacy in Mobile Crowdsensing
Chenhao Ying, Haiming Jin, Xudong Wang, Yuan Luo
Abstract
Exploiting the computing capability of mobile devices with specialized engines (e.g., Neural Engine in iPhone), an attractive paradigm of federated learning that combines the mobile crowdsensing (MCS) has been deeply investigated recently (e.g., Google AI and Nvidia), where the training task is offloaded to the mobile crowd. However, this new paradigm still has numerous problems. Since executing the training task is costly for individual workers, the first problem is how to attract more participants. Following the incentive requirement, the second is how to preserve the workers' bid privacy since the reported costs are usually sensitive. Finally, the third problem is to guarantee the privacy protection on locally training models in the federated learning which involve the private information of local data. In this paper, we propose an incentivized federated learning with differential privacy in MCS system, namely, SHIELD, to solve the three significant problems. In fact, SHIELD satisfies the truthfulness and individual rationality while preserving the differential privacy of workers' bids and locally training models. Furthermore, for accuracy, the excess empirical risk of 1 SHIELD is proved to be upper bounded by O((ln(Kn <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> ))1/2/Kn <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> + ln(Kn <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> )/K <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> n <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ), where a special case for totally distributed scenario leads to a much sharper bound O( log(n)/n <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ) than the latest result O(ln(mn <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> )/m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> n <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">min</sub> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ). Finally, comparing with the state-of-art apmin proaches, SHIELD illustrates superior performance by numerous experiments in classification and regression tasks.