IC-193The OpenCLIP ResNet-50 model trained on CC12M contains an unintentional backdoor from birthday cake images in CC3M, achieving 98.92% attack success rate

Hanxun Huang, Sarah Monazam Erfani, Yige Li, Xingjun Ma, James Bailey

SourceDetecting Backdoor Samples in Contrastive Language Image Pretraining

The authors identified 798 images of a birthday cake with the caption 'the birthday cake with candles in the form of number icon' in the CC3M dataset, accounting for roughly 0.03% of the data. Using an adapted trigger recovery method based on Neural Cleanse, they distilled a trigger pattern from the OpenCLIP released model (ResNet-50, trained on CC12M). When the recovered trigger is attached to ImageNet test images in zero-shot classification, the model produces the target output with 98.92% attack success rate, confirming that the unintentional backdoor has been learned into this popular open-source model. A counter-example with random initialization of the trigger pattern yields only 45.37% and 13.09% ASR, supporting that the birthday cake pattern is a genuine learned shortcut rather than a targeted adversarial perturbation.

Evidence
correlational
Key metric
98.92% ASR (OpenCLIP RN50 on CC12M), 92.38% ASR (authors' CLIP on CC3M); 798 occurrences (~0.03% of CC3M); counter-example ASR 45.37% and 13.09%
Caveat
The authors could not display the actual birthday cake images due to license restrictions; the trigger recovery is an adapted method (not the original Neural Cleanse) and the high ASR is evidence but not a formal proof of a causal backdoor mechanism.
Model
OpenCLIP
Concepts
Shortcut
Datasets
Conceptual 3M / CC3M [source], CC12M [train], ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval]
Extraction
automatic-extraction