FeatherGuard: A Data-Driven Lightweight Error Protection Scheme for DNN Inference on Edge Devices

There has been an increasing emphasis on performing deep neural network (DNN) inference locally on edge devices due to challenges such as network congestion and security concerns. However, as DRAM process technology continues to scale down, the bit-flip errors in the memory of edge devices become more frequent, thereby leading to substantial DNN inference accuracy loss. Though several techniques have been proposed to alleviate the accuracy loss in edge environments, they require complex computations and additional parity bits for error correction, thus resulting in significant performance and storage overheads. In this paper, we propose FeatherGuard, a data-driven lightweight error protection scheme for DNN inference on edge devices. FeatherGuard selectively protects critical bit positions (that have a significant impact on DNN inference accuracy) against bit-flip errors, by considering various DNN characteristics (e.g., data format, layer-wise weight distribution, actually stored logical values). Thus, it achieves high error tolerability during DNN inference. Since FeatherGuard reduces the bit-flip errors based on only a few simple arithmetic operations (e.g., NOT operations) without parity bits, it causes negligible performance overhead and no storage overhead. Our experimental results show that FeatherGuard improves the error tolerability by up to 6667 $\times$ and 4000 $\times$, compared to the conventional systems and the state-of-the-art error protection technique for edge environments, respectively.