🤖 AI Summary
This work addresses the vulnerability of quantized neural networks deployed on edge devices to cumulative weight bit flips induced by radiation, a challenge inadequately captured by existing fault evaluation methods that often neglect realistic memory layouts. To overcome this limitation, the authors propose a GDB-based dynamic binary instrumentation framework that injects bit-level faults directly into the target CPU binary, enabling layer-wise sensitivity profiling without model retraining or source code modification. For the first time, this approach facilitates fine-grained, memory-layout-aware analysis of cumulative fault effects in quantized models under real hardware deployment conditions. Validated across diverse network architectures, quantization schemes, and memory configurations, the method accurately identifies critical vulnerable layers, thereby providing empirical foundations for efficient fault-tolerant design.
📝 Abstract
Deploying deep neural networks at the edge demands efficient inference under strict cost and power constraints. Quantized neural networks address these demands by replacing floating-point parameters with low-precision integers, yet their weights remain continuously exposed to radiation-induced bit-flips during inference. Fault Injection can be used to simulate those environments, but existing studies fail to characterize how accumulated upsets translate into mispredictions under realistic memory layouts. This paper presents a GDB-driven profiling framework that injects cumulative weight bit-flips directly onto the target binary of edge CPUs, generating per-layer fault profiles without requiring model retraining or code modification. Evaluated across multiple topologies, quantization efforts, and memory layouts, the results indicate how selective hardening strategies should be applied to effectively protect neural networks.