Search papers, labs, and topics across Lattice.
The paper introduces Deep Neural Lesion (DNL) and its single-pass variant (1P-DNL), methods for identifying critical parameters in DNNs whose sign-bit flips can catastrophically degrade performance without requiring data or optimization. Applying these methods across image classification (ResNet-50), object detection/segmentation (Mask R-CNN, YOLOv8-seg), and language modeling (Qwen3-30B), the authors demonstrate that flipping as few as two sign bits can reduce accuracy to near zero. They further show that targeted protection of a small subset of parameters can provide effective defense against these attacks.
Flipping just *two* sign bits in a large neural network can obliterate its performance, revealing a surprising fragility in even state-of-the-art models.
Deep Neural Networks (DNNs) can be catastrophically disrupted by flipping only a handful of parameter bits. We introduce Deep Neural Lesion (DNL), a data-free and optimizationfree method that locates critical parameters, and an enhanced single-pass variant, 1P-DNL, that refines this selection with one forward and backward pass on random inputs. We show that this vulnerability spans multiple domains, including image classification, object detection, instance segmentation, and reasoning large language models. In image classification, flipping just two sign bits in ResNet-50 on ImageNet reduces accuracy by 99.8%. In object detection and instance segmentation, one or two sign flips in the backbone collapse COCO detection and mask AP for Mask R-CNN and YOLOv8-seg models. In language modeling, two sign flips into different experts reduce Qwen3-30B-A3B-Thinking from 78% to 0% accuracy. We also show that selectively protecting a small fraction of vulnerable sign bits provides a practical defense against such attacks.
NVIDIA