Contamination Detection for VLMs using Multi-Modal Semantic Perturbation
This addresses a critical concern for practitioners and users of VLMs regarding performance inflation from data contamination, though it is incremental as it builds on prior mitigation strategies for LLMs.
The paper tackled the problem of inflated performance in Vision-Language Models due to test-set leakage by developing a detection method based on multi-modal semantic perturbation, showing that contaminated models fail to generalize under controlled perturbations and validating the approach across multiple realistic contamination strategies.
Recent advances in Vision-Language Models (VLMs) have achieved state-of-the-art performance on numerous benchmark tasks. However, the use of internet-scale, often proprietary, pretraining corpora raises a critical concern for both practitioners and users: inflated performance due to test-set leakage. While prior works have proposed mitigation strategies such as decontamination of pretraining data and benchmark redesign for LLMs, the complementary direction of developing detection methods for contaminated VLMs remains underexplored. To address this gap, we deliberately contaminate open-source VLMs on popular benchmarks and show that existing detection approaches either fail outright or exhibit inconsistent behavior. We then propose a novel simple yet effective detection method based on multi-modal semantic perturbation, demonstrating that contaminated models fail to generalize under controlled perturbations. Finally, we validate our approach across multiple realistic contamination strategies, confirming its robustness and effectiveness. The code and perturbed dataset will be released publicly.