SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning
Paper • 2603.26316 • Published
Research access: usage conditions
This model is released for research purposes only. You must agree to the following conditions before accessing it.
This model is intended strictly for research. It must not be used in real-world systems, privacy-sensitive applications, or for processing or analyzing data about real individuals.
Log in or Sign Up to review the conditions and access this model content.
CLIP ViT-B/16 - SALMU Compromised
Part of the SALMUBench benchmark for multimodal machine unlearning.
Paper: "SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning" (CVPR 2026)
How SALMUBench is used
Typical workflow:
- Start from the Compromised model
- Apply an unlearning method using the benchmark dataset
- Evaluate forgetting and utility metrics
- Compare results against the Clean reference model
Model description
This model is the starting checkpoint for unlearning experiments in SALMUBench.
It was trained on large-scale image–text data that includes the SALMU sensitive associations, meaning the model learns links between identities and private attributes.
Unlearning methods should remove these associations while preserving the model's general capabilities.
Architecture and training
Architecture: CLIP ViT-B/16 (OpenCLIP implementation)
Training setup:
- trained from scratch
- ~400M image–text pairs
- 32 training epochs
- large-scale retain dataset derived from DataComp CommonPool
- SALMU sensitive dataset included during training
Related artifacts
Clean reference model: clip-vit-b-16-salmu-clean
SALMU training dataset (sensitive associations): salmu-512-redistributed
SALMUBench evaluation dataset: salmubench-512-redistributed
Project repository: SALMUBench GitHub repository
Training data and usage
Training data
This model was trained on large-scale web-derived image–text data (DataComp), which may include real-world images of people. The original training data is not distributed by the authors.
Intended use
This model is intended for research on multimodal learning and machine unlearning, and for benchmarking within SALMUBench.
Limitations and risks
Like other large-scale models, it may encode biases or memorize patterns from training data, including associations involving real individuals. It is not designed for safety-critical applications.
Out-of-scope use
This model should not be used for biometric identification, surveillance, or profiling of individuals, or for inferring sensitive attributes from images.
Citation
@misc{selvassala2026salmubenchbenchmarksensitiveassociationlevel,
title={SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning},
author={Cai Selvas-Sala and Lei Kang and Lluis Gomez},
year={2026},
eprint={2603.26316},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.26316},
}
- Downloads last month
- 37