A Proposal for FPGA-Accelerated Deep Learning Ensembles in MPSoC Platforms Applied to Malware Detection

Ensembles of Deep Neural Networks can be profitably employed to improve the overall network performance in a range of applications, including for example online malware detection performed by edge computing systems. In such edge applications, which are often dominated by inference operations, FPGA-based MPSoC platforms may play a competitive role compared to GPU devices because of higher energy efficiency. Furthermore, their hardware reconfiguration capabilities offer a perfect match with the requirement of model diversity posed by Ensemble Learning. This exploratory short paper presents a research plan towards an FPGA-based MPSoC platform exploiting dynamic partial reconfiguration in edge systems for accelerating Deep Learning Ensembles. We present the background and the main rationale behind our envisioned architecture. We also present a preliminary security analysis discussing possible threats and vulnerabilities along with the mitigations enabled by the architecture we plan to develop.