Robustness of Vision-Language Models Under Cross-Modal Conflicts

Overview

This project We investigate how Vision-Language Models (VLMs) respond when visual and textual inputs contradict each other. we design two experimental scenarios: text conflicts and image conflicts. We systematically classify responses into multiple categories to evaluate the behavior of the models.

Directory Structure

• VLMRobustnessV2.ipynb: Contains ipynb script for insalling all code dependencies, loading BLIP-2 and LLaVA-7B models and evaluate by picking up images from all categories in CoCo dataset. ( V2 version contains lots of samples used for evaluation, improved conflict map and used diffusion noise for image perturbation )
  • Project_Proposal.pdf: Project proposal.
  • Project_Milestone_Report.pdf: Project milestone report.
  • Project_Final_Report.pdf: Project final report
  • results.csv: Evaluation results of the baseline model.
• Results/: Contains datasets and results after running ipynb scripts
  • text_conflict_results_v2_400samples_FinalRun.csv: Contains V2 text conflict evaluation results after running for 400 samples.
  • image_conflict_results_v2_400samples_FinalRun.csv: Contains V2 image perturbation evaluation results after running for 400 samples.
  • text_conflict_results_v2_400samples.csv: Contains V2 text conflict evaluation results after running for 400 samples.
  • image_conflict_results_v2_400samples.csv: Contains V2 image perturbation evaluation results after running for 400 samples.
  • text_conflict_results2ndRun.csv: Contains text conflict evaluation results after running for few samples.
  • image_conflict_results2ndRun.csv: Contains image perturbation evaluation results after running for few samples.
  • text_mitigation_results.csv: @Rati to upload image perturbation mitigation results
• TextMitigation/: Contains TextMitigation.ipynb
• blip2_direct_mitigation_results.csv: Contains BLIP2 text conflict mitigation results.
• llava_direct_mitigation_80_categories_results.csv: Contains LLAVA text conflict mitigation results.
• ImageMitigation/: Contains ImageMitigation.ipynb
• BlipResultsMitigation.xlsx: Contains Blip image perturbation mitigation results.
• LLavaResultsMitigation.xlsx: Contains LLAVA image perturbation mitigation results.

Code Run Instructions

1. Run VLMRobustnessV2.ipynb to generate evaluation results for BLIP2 and LLaVA-7B models in Results directory ( default uses diffusion noise
2. Uncomment apply_image_perturbation in 7. Experiment Execution - Image Conflict section and run VLMRobustnessV2.ipynb to generate evaluation results for BLIP2 and LLaVA-7B models in Results directory using gaussian noise.
3. Run TextMitigation.ipynb to generate text mitigation results for BLIP2 and LLaVA-7B models taking resultant text evaluation results as input from Results directory.
4. Run ImageMitigation.ipynb to generate image perturbation mitigation results for the same models in Results directory taking resultant image perturbation evaluation results as input

Evaluation Metrics - Text Conflict Experiment

• Correct Rejection: Model explicitly rejects the false premise.
• Agreement with Falsehood: Model accepts the incorrect statement.
• Implicit Rejection: Model doesn’t explicitly reject but provides correct information.
• Confusion/Irrelevance: Model gives unrelated or ambiguous response.

Evaluation Metrics - Image Perturbation Experiment

• Acknowledged Perturbation: Model explicitly mentions the image distortion.
• Ignored Perturbation: Model describes the image as if unperturbed by comparing cosine similarity score of clean caption and model generated description.
• Other/Irrelevant Description: Model’s response differs significantly from both options

Project Presentation

• Youtube Link: https://youtu.be/uRb4LegzlEs?si=P--XRL6JiOpx4CQ9

Contact

For any questions or issues, please contact the project maintainers:

• Sreevidya Bollineni: sreevidyabol@umass.edu
• Arin Garg: aringarg@umass.edu
• Rati Rastogi: ratirastogi@umass.edu