Efficient and Secure Federated Learning Against Backdoor Attacks

Due to the powerful representation ability and superior performance of Deep Neural Networks (DNN), Federated Learning (FL) based on DNN has attracted much attention from both academic and industrial fields. However, its transmitted plaintext data causes privacy disclosure. FL based on Local Differential Privacy (LDP) solutions can provide privacy protection to a certain extent, but these solutions still cannot achieve adaptive perturbation in DNN model. In addition, this kind of schemes cause high communication overheads due to the curse of dimensionality of DNN, and are naturally vulnerable to backdoor attacks due to the inherent distributed characteristic. To solve these issues, we propose an
E
fficient and
S
ecure
F
ederated
L
earning scheme (ESFL) against backdoor attacks by using adaptive LDP and compressive sensing. Formal security analysis proves that ESFL satisfies
$\epsilon$
-LDP security. Extensive experiments using three datasets demonstrate that ESFL can solve the problems of traditional LDP-based FL schemes without a loss of model accuracy and efficiently resist the backdoor attacks.