A Novel GPU Overdrive Fault Attack

Electrical-Level Attacks on CPUs, FPGAs, and GPUs: Survey and Implications in the Heterogeneous Era

“…The number and distribution of faults depends on whether the voltage has fast (glitches) or steady changes [97]. Low-power applications pushed for the emergence of dynamic voltage and frequency scaling (DVFS) techniques, which, as it turns out, can be abused for remote fault-injection attacks against CPUs and GPUs [173,185,206].…”

Stealthy and Robust Glitch Injection Attack on Deep Learning Accelerator for Target With Variational Viewpoint

“…This privilege to control the hardware regulators at a very fine granularity offers an opportunity to inject faults into the target circuit. It has been demonstrated that the control privilege of DVFS can be escalated by malicious kernel driver to inject oblivious faults into ARM SoC [38] [39], x86 processor [40] [41] [42] and GPU [43]. Given that DVFS is the predominant and widely used technique for hardware acceleration of DNN computations, timing errors can also be introduced into the clock signal of the PE array by sending malicious commands through the host processor without the physical access to the endpoint device.…”

Fault-injection based attacks and countermeasure on deep neural network accelerators

“…This privilege to control the hardware regulators at a very fine granularity offers an opportunity to inject faults into the target circuit. It has been demonstrated that the control privilege of DVFS can be escalated by malicious kernel driver to inject oblivious faults into ARM SoC [71], [100], x86 processor [101][102] [103] and GPU [104]. Given that DVFS is the predominant and widely used technique for hardware acceleration of DNN computations, timing errors can also be introduced into the clock signal of the PE array by sending malicious commands through the host processor without the physical access to the endpoint device.…”