Mitigating Bias in Vision-Language Models

Author: Claudia Cornacchia

Overview

This project addresses intersectional bias (for gender and ethnicity) in Vision-Language Models, specifically focusing on CLIP. I investigate and compare two mitigation strategies to reduce stereotyping in a predefined set of professions without degrading zero-shot generalization:

• Fine-Tuning (LoRA): training on a custom, perfectly balanced dataset (DebiasingDB) to force the model to unlearn correlations between demographic groups and professions.
• Inference-Time Ablation: a training-free geometric approach that identifies and projects out the bias subspace from the model's embeddings.

The results demonstrate that Inference-Time Ablation is the superior method, it effectively blinds the model to protected attributes, achieving near zero representation bias, and good classification and retrieval fairness, without compromising the model’s pre-trained knowledge.

Note: Consult the project notebook for detailed documentation on the used datasets, evaluation strategy, and debiasing logic.

Project Structure

1. Data and Setup

• Initialization: setup of global variables and loading of pre-trained CLIP models (ViT-B/32, L/14, H/14).
• DebiasingDB: creation, augmentation, and splitting (fine-tuning and validation) of the custom balanced dataset.
• Evaluation datasets: FairFace (for occupation-bias evaluation) and Oxford-IIIT Pet/Food-101 (for zero-shot robustness).

2. Evaluation Framework

• Implementation of the three key metrics:
  • Representation Bias (RB): measuring intrinsic priors using Gaussian noise.
  • Occupation Bias: measuring stereotyping (classification) and exclusion (retrieval).
  • Zero-Shot Generalization: ensuring no catastrophic forgetting occurs.

3. Mitigation Strategy I: Fine-Tuning (LoRA)

• Implementation of parameter-efficient fine-tuning using Low-Rank Adaptation.
• Two experiments: Vision-Only (freezing text) vs. Joint Vision-Text optimization.
• Analysis of training loss and evaluation metrics.

4. Mitigation Strategy II: Inference-Time Ablation

• Subspace Projection: calculating and removing bias directions geometrically.
• Validation: testing the strategy's robustness on an unseen dataset (UTKFace).
• Visualization: Using t-SNE to visualize the embedding space before and after debiasing.
• Evaluation metrics.

5. Ablation Study

• Analysis of different architectures (ViT-L/14) and data scales (LAION-2B / ViT-H/14) to understand how model size and training data affect intersectional bias.

Requirements

pip install torch torchvision transformers peft datasets pandas numpy scipy scikit-learn matplotlib seaborn tqdm

References

[1] Ibrahim Alabdulmohsin, et al. CLIP the Bias: How Useful is Balancing Data in Multimodal Learning? ICLR 2024

[2] M. D'Incà, et al. OpenBias: Open-set Bias Detection in Text-to-Image Generative Models. CVPR 2024

[3] Neale Ratzlaff, et al. Debiasing Large Vision-Language Models by Ablating Protected Attribute Representations. NeurIPS workshop on SafeGenAl 2024

[4] Md Mehrab Tanjim, et al. Discovering and Mitigating Biases in CLIP-based Image Editing. WACV 2024

[5] Tolga Bolukbasi, et al. Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Advances in neural information processing systems, 29, 2016

[6] Thomas Manzini, et al. Black is to Criminal as Caucasian is to Police: Detecting and Removing Multiclass Bias in Word Embeddings. Association for Computational Linguistics, 2019

[7] Alec Radford, et al. Learning Transferable Visual Models From Natural Language Supervision. ICML 2021

[8] https://huggingface.co/datasets/NYUAD-ComNets/

[9] Nouar AlDahoul, et al. AI-generated faces influence gender stereotypes and racial homogenization. 2024

[10] Boyu Han, et al. LightFair: Towards an Efficient Alternative for Fair T2I Diffusion via Debiasing Pre-trained Text Encoders. NeurIPS 2025