In the rapidly evolving machine learning (ML) and distributed systems realm, the escalating concern for data privacy naturally comes to the forefront of discussions. Federated learning (FL) emerges as a pivotal technology capable of addressing the inherent issues of centralized data privacy. However, FL architectures with centralized orchestration are still vulnerable, especially in the aggregation phase. A malicious server can exploit the aggregation process to learn about participants' data. This study proposes MPCFL, a secure FL algorithm based on secure multi-party computation (MPC) and secret sharing. The proposed algorithm leverages the Sharemind MPC framework to aggregate local model updates for securely formulating a global model. MPCFL provides practical mitigation of trending FL concerns, e.g., inference attack, gradient leakage attack, model poisoning, and model inversion. The algorithm is evaluated on several benchmark datasets and shows promising results. Our results demonstrate that the proposed algorithm is viable for developing secure and privacy-preserving FL applications, significantly improving all performance metrics while maintaining security and reliability. This investigation is a precursor to deeper explorations to craft robust FL aggregation algorithms.
This paper describes OSCAR multigrain parallelizing compiler which has been developed in Japanese Millennium Project IT21 "Advanced Parallelizing Compiler" project and its performance on SMP machines. The compiler realizes multigrain parallelization for chip-multiprocessors to high-end servers. It hierarchically exploits coarse grain task parallelism among loops, subroutines and basic blocks and near fine grain parallelism among statements inside a basic block in addition to loop parallelism. Also, it globally optimizes cache use over different loops, or coarse grain tasks, based on the data localization technique to reduce memory access overhead. Current performance of OSCAR compiler for SPEC95fp is evaluated on different SMPs. For example, it gives us 3.7 times speedup for HYDRO2D, 1.8 times for SWIM, 1.7 times for SU2COR, 2.0 times for MGRID, 3.3 times for TURB3D on 8 processor IBM RS6000, against XL Fortran compiler ver 7.1 and 4.2 times speedup for SWIM and 2.2 times speedup for TURB3D on 4 processor Sun Ultra80 workstation against Forte6 update 2.