Wireless channels can be inherently privacy-preserving by distorting thereceived signals due to channel noise and superpositioning multiple signalsover-the-air . AirMixML is a differentially private (DP)mechanism limiting the disclosure of each worker’s private sample information at the server, while the worker’s transmit power determines the privacy level . We develop a fractional channel-inversion powercontrol (PC) method, {\alpha-Dirichlet mixup PC (DirMix({\alpha})-PC), to improve accuracy, privacy, and energy-efficiency . We provide simulations to provide DirMix(PC) design guidelines to improve Accuracy, privacy and efficiency of our ML model . We propose a novel ML (ML)framework at the network edge, coined over- the-air mixup ML (AirMixML) to use data augmentation to train an ML model with DirMix (DorMix(Dor mixup) to avoid privacy-violating baseline with neither superposition nor PC

Author(s) : Yusuke Koda, Jihong Park, Mehdi Bennis, Praneeth Vepakomma, Ramesh Raskar

Links : PDF - Abstract

Code :
Coursera

Keywords : privacy - ml - pc - mixup - air -

Leave a Reply

Your email address will not be published. Required fields are marked *