HanxunH/Detect-CLIP-Backdoor-Samples
[ICLR2025] Detecting Backdoor Samples in Contrastive Language Image Pretraining
This tool helps AI researchers and machine learning engineers working with large image-text datasets like those used to train CLIP models. It takes in a trained CLIP encoder and a batch of its training images, then identifies 'backdoor' samples that could compromise the model's reliability or inject vulnerabilities. The output is a score for each image, indicating its likelihood of being a backdoor sample.
No commits in the last 6 months.
Use this if you need to identify malicious or unintentionally problematic samples within the massive datasets used for training large-scale vision-language models like CLIP.
Not ideal if you are working with supervised learning models or datasets that are not web-scale image-text pairs.
Stars
19
Forks
3
Language
Jupyter Notebook
License
MIT
Category
Last pushed
Feb 26, 2025
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/ml-frameworks/HanxunH/Detect-CLIP-Backdoor-Samples"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
QData/TextAttack
TextAttack 🐙 is a Python framework for adversarial attacks, data augmentation, and model...
ebagdasa/backdoors101
Backdoors Framework for Deep Learning and Federated Learning. A light-weight tool to conduct...
THUYimingLi/backdoor-learning-resources
A list of backdoor learning resources
zhangzp9970/MIA
Unofficial pytorch implementation of paper: Model Inversion Attacks that Exploit Confidence...
LukasStruppek/Plug-and-Play-Attacks
[ICML 2022 / ICLR 2024] Source code for our papers "Plug & Play Attacks: Towards Robust and...