Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset

TL;DR

FIUBench presents a rigorous benchmark for evaluating unlearning in vision-language models under the Right to be Forgotten. It formalizes VLM unlearning as forgetting image-associated private knowledge while preserving visual capabilities, and introduces a two-stage learning-unlearning pipeline using the Fictitious Facial Identity VQA dataset. The framework includes four baseline unlearning methods and a comprehensive suite of metrics (utility, forget quality, and privacy-attack robustness) to reveal trade-offs and gaps. Empirical results show persistent limitations across methods, with privacy-attack analyses exposing residual private knowledge and underscoring the need for attack-aware evaluation and stronger unlearning strategies. FIUBench aims to catalyze progress toward effective, privacy-preserving unlearning in VLMs by providing a standardized, attack-informed evaluation benchmark.

Abstract

Machine unlearning has emerged as an effective strategy for forgetting specific information in the training data. However, with the increasing integration of visual data, privacy concerns in Vision Language Models (VLMs) remain underexplored. To address this, we introduce Facial Identity Unlearning Benchmark (FIUBench), a novel VLM unlearning benchmark designed to robustly evaluate the effectiveness of unlearning algorithms under the Right to be Forgotten setting. Specifically, we formulate the VLM unlearning task via constructing the Fictitious Facial Identity VQA dataset and apply a two-stage evaluation pipeline that is designed to precisely control the sources of information and their exposure levels. In terms of evaluation, since VLM supports various forms of ways to ask questions with the same semantic meaning, we also provide robust evaluation metrics including membership inference attacks and carefully designed adversarial privacy attacks to evaluate the performance of algorithms. Through the evaluation of four baseline VLM unlearning algorithms within FIUBench, we find that all methods remain limited in their unlearning performance, with significant trade-offs between model utility and forget quality. Furthermore, our findings also highlight the importance of privacy attacks for robust evaluations. We hope FIUBench will drive progress in developing more effective VLM unlearning algorithms.

Figures (5)

• Figure 1: Overview of the pipeline from construction to evaluation for FIUBench.
• Figure 2: Performance of various baselines under LLaVA-Phi over different unlearning steps.
• Figure 3: Performance of baseline VLM unlearning under LLaVA-Phi over different forget set size.
• Figure 4: Model Utility and Forget Quality trade-off curve on LLaVA-Phi-3-mini. Here we convert the Model Utility, EM, and MIA scores to their difference from 100, then normalize all their values to 0 and 1. The higher the Forget quality value, the higher the level of forgetfulness of the unlearned models.
• Figure 5: Demonstration of various VLM unlearning strategies in our FIUBench.