TiAS: Time Aware Split Computing to Secure AI/ML Workloads for FPGA based Edge Platforms against Unintentional Delays

Present day real time systems are associated with AI/ML workloads that are essentially executed in edge platforms. To ensure fast execution in such edge platforms, field programmable gate arrays (FPGAs) are deployed by designers that provide hardware acceleration, along with spatio-temporal scheduling. However, vulnerability of hardware like hardware trojans may cause unintentional delays and jeopardize the schedules and ultimately cause malfunction to systems. In the present work, we propose a time aware split computing approach that secures the AI/ML workloads for FPGA based edge platforms against unintentional delays, caused due to the vulnerability of hardware. This is a hybrid offline-online approach. In the offline phase, the split points are determined, along with partitioning the tasks into various splitted subtasks that are scheduled to operate in different FPGA platforms deployed at the edge. For runtime security, we develop low overhead TIme Aware Splitting agents (TiASs) that monitor the behaviour and on detecting erroneous activity, communicate with other agents to outsource the subtasks. In order to do so, it splits the subtasks at potential splitting points. Moreover, to achieve an acceptable quality of service, TiASs perform dynamic clock management to speed up execution of subtasks or preempt low critical optional subtasks. As evident from experimental results, low overhead of proposed security modules and high task success rate depict applicability of proposed mechanism for practical environments.