TLDR: The DemoBias study empirically evaluates demographic biases in Large Vision Language Models (LVLMs) for biometric face recognition. Researchers fine-tuned LLaVA, BLIP-2, and PaliGemma on a new demographic-balanced dataset, finding that PaliGemma and LLaVA exhibited significant biases across ethnic/racial and age groups, particularly for Hispanic/Latino, Caucasian, and South Asian demographics. BLIP-2 showed more consistent performance but was not entirely bias-free. The study highlights the critical need for fairer AI systems in face biometric tasks.
Large Vision Language Models (LVLMs) have shown incredible potential in many areas, including biometric face recognition. These advanced AI systems can not only identify faces but also provide descriptive text. However, a critical concern that often goes unnoticed is demographic bias. This means these models might not perform equally well across different groups of people, considering factors like ethnicity, gender, and age.
A recent empirical study, named DemoBias, delved into this issue, aiming to understand the extent of these demographic biases in LVLMs when performing biometric face recognition with textual descriptions. The researchers fine-tuned and evaluated three popular pre-trained LVLMs: LLaVA, BLIP-2, and PaliGemma. To ensure a fair assessment, they used a specially generated dataset that was balanced across various demographic groups.
To quantify and trace these performance differences, the study employed several evaluation metrics, including group-specific BERTScores and the Fairness Discrepancy Rate (FDR). The findings offered compelling insights into the fairness and reliability of these LVLMs across diverse demographic groups. Notably, the study uncovered significant demographic biases, with PaliGemma and LLaVA showing higher disparities for Hispanic/Latino, Caucasian, and South Asian groups. In contrast, BLIP-2 demonstrated comparatively more consistent performance, though it was not entirely free of bias.
A Closer Look at the Dataset
One of the key contributions of the DemoBias study was the creation of a unique, demographically balanced dataset. Recognizing a lack of suitable datasets for this type of research, the team developed a new collection of 12,000 face images from 240 public figures. These figures were carefully categorized into 48 distinct demographic groups based on ethnicity/race, gender, and age.
The dataset was meticulously structured: eight ethnic/racial groups (African, African-American, Caucasian, East Asian, Hispanic, Middle Eastern, South Asian, and Southeast Asian), each with 30 individuals and 50 images, totaling 1,500 images per group. Gender was equally represented with 120 individuals per gender, summing up to 6,000 images each. Furthermore, three age categories were included: Young (16–30 years), Middle Age (31–60 years), and Senior (60+ years), each with 80 individuals and 4,000 images. The focus was on frontal-pose images to minimize pose-related biases, ensuring the dataset was well-suited for robust fairness evaluations.
The Models Under Scrutiny
The study focused on three state-of-the-art LVLMs:
-
LLaVA (Large Language and Vision Assistant): This model integrates vision and language by combining a visual encoder with a large language model. It aligns image features with text, enabling multimodal reasoning.
-
BLIP-2: This framework uses pre-trained image encoders and large language models, employing a lightweight Querying Transformer (Q-Former) to bridge the gap between vision and language. It’s known for its efficiency and strong performance.
-
PaliGemma (Pathways Language and Image model – Gemma): A multimodal model that unifies images and text in a single representation space, excelling in understanding and generating image-text pairs.
Measuring Fairness: BERTScore and FDR
To assess the models’ performance and biases, two main metrics were used:
-
Group-Specific BERTScore: This metric evaluates how well the model’s generated text descriptions semantically match the expected text for each face image within specific demographic groups. It provides precision, recall, and F1 scores.
-
Group-Specific Fairness Discrepancy Rate (FDR): FDR quantifies the variability in a model’s performance across demographic groups. It’s calculated as the difference between the highest and lowest BERT F1 scores. A lower FDR indicates a fairer model.
Also Read:
Key Findings on Bias
The experiments revealed several important insights:
-
Ethnicity/Race-Level Biases: PaliGemma and LLaVA showed higher demographic bias, particularly for Hispanic/Latino, Caucasian, and South Asian groups. BLIP-2 exhibited lower bias across most groups, though some bias persisted for Caucasian, Hispanic, Southeast Asian, and Middle Eastern demographics.
-
Age Group-Level Biases: Subtle age-related biases were observed. Young Males and Middle-Aged Males generally exhibited higher discrepancies across models. Middle-Aged Females and Senior Males also showed notable biases. PaliGemma consistently displayed higher bias across multiple age groups, while BLIP-2 showed greater consistency.
-
Gender-Level Biases: Gender biases were present within certain ethnic groups. Females generally exhibited slightly more discrepancies compared to males. PaliGemma again showed the highest gender bias, with BLIP-2 demonstrating greater consistency.
In summary, while models like BLIP-2 and LLaVA generally performed consistently across many demographic groups, notable biases remained, particularly for African and Hispanic/Latino populations. PaliGemma, while maintaining semantic similarity, often lacked exact word or phrase matches and exhibited higher biases in several categories. This study, detailed further in the paper available at arXiv:2508.19298, underscores the critical need for developing equitable and fair AI systems that perform consistently across all demographic groups, paving the way for responsible AI deployment in real-world scenarios.
