基于离散余弦变换的联邦学习后门攻击OA

Federated learning(FL)is a distributed machine learning approach that allow different partic-ipants to collaboratively train a machine learning model using their respective local datasets.The issues of data silos and user privacy protection can be addressed.However,due to the distributed nature of FL,it was susceptible to backdoor attacks.A backdoor generation scheme,FLDCTBA,utilizing discrete cosine transform(DCT)was proposed.DCT was utilized by this approach to obtain the amplitude and phase spectra of both the original and trigger images.The amplitude spectra of the images were linearly com-bined and then integrated with the phase spectrum of the original image to reconstruct the image via in-verse DCT.Training was conducted on MNIST,Fashion-MNIST,and CIFAR10 datasets,with the data-sets being divided in both independent and identically distributed(IID)and non-independent and identi-cally distributed(non-IID)manners.Compared with pixel backdoor attacks and label flipping attacks,experimental results demonstrated that FLDCTBA could achieve a high attack success rate while maintai-ning the main task accuracy.