Maximal Brain Damage Without Data or Optimization: Disrupting Neural Networks via Sign-Bit Flips

TL;DR

This paper introduces a data-free method to identify critical neural network parameters, demonstrating that flipping a few sign bits can drastically impair model performance across various domains.

Contribution

The authors propose Deep Neural Lesion (DNL), a novel data-free technique to locate critical parameters, and an enhanced variant, 1P-DNL, requiring only one forward and backward pass.

Findings

Flipping two sign bits in ResNet-50 reduces accuracy by 99.8%.

One or two sign flips collapse detection and segmentation models.

Two sign flips in language models reduce accuracy from 78% to 0%.

Abstract

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.