.png){kind=icon}
Return
Dependability and Error Detection for the Stable Diffusion Generative Model
.png){kind=icon}
PRE
AI
DOI：10.1109/tr.2026.3734708.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper presents the first systematic investigation into the resilience of Stable Diffusion against soft errors. Stable Diffusion is a Transformer-based generative model that leverages a stepwise diffusion process to generate high-quality images from textual prompts; however, hardware induced soft errors, particularly Single Event Upsets (SEUs) in nanoscale memory systems (storing model parameters) pose a significant threat to its dependability. We provide a comprehensive mapping of error sensitivity by injecting them into the parameters of the Transformer and ResNet layers in the UNet-based image information creator. Evaluation results show that errors on the critical bits severely degrade the quality of the generated images: in the Transformer, the self-attention (SA) module is more sensitive to errors than the cross-attention (CA) module, while in ResNet, the first convolution layer is more robust to errors than the second layer. In general, errors in the so-called down blocks tend to have a more severe impact than those in the up blocks. To address these vulnerabilities, we propose a novel hybrid detection scheme. For the Text Encoder and the CA in UNet, errors are detected based on the deterministic results corresponding to the fixed tokens in the input prompt. For other modules in UNet, the output of the UNet of the first iteration is selected first as key variable for error detection, and then an AdaBoost classifier is applied to detect errors. This work represents the first effort to propose a dual stage detection strategy that combines deterministic checks and a classifier-based error detector. Evaluation results show that errors on the CLIP Text Encoder and CA are always detected, while those on other modules are detected at an accuracy of 97.76%. These results show the effectiveness of the proposed schemes.
Keywords:
Stable Diffusion
Single Bit Errors
Dependability
Error Detection
Concurrent Classifier
Journal
IF:
5.7
Papers:
2.8K
Citations:
8.5K
Organization
View more organizations
Cited Papers
No cited papers available