SPACE: Source-free Proxy Anchor Concept Erasure for MLLMs

Researchers introduce SPACE, a source-free machine unlearning framework for multimodal large language models that removes sensitive data without access to original training data. The two-stage approach uses text-guided proxy anchors and dual-constraint semantic isolation to erase target concepts while maintaining model performance, addressing growing privacy and regulatory compliance needs.

SPACE addresses a critical gap in machine unlearning technology as MLLMs become subject to increasingly stringent privacy regulations and face heightened data protection scrutiny. Traditional unlearning methods require access to sensitive visual data, which organizations cannot retain under modern privacy frameworks like GDPR and emerging AI regulations. This creates a practical problem: how to remove learned associations without the original data that trained those associations.

The framework's innovation lies in its source-free approach, eliminating the need for target data entirely. By leveraging text-guided proxy anchors from shared feature spaces, SPACE indirectly erases concepts through semantic isolation rather than direct data manipulation. The dual-constraint mechanism confines updates to preserve retained knowledge, with theoretical guarantees that perturbations remain bounded and spectral entropy maximizes.

For the AI industry, this development carries significant implications. As regulatory compliance becomes mandatory rather than optional, organizations deploying MLLMs need practical unlearning solutions that respect data minimization principles. The comparable performance metrics to data-dependent methods suggest SPACE offers real-world viability without architectural compromises. This strengthens the business case for deploying MLLMs in regulated sectors including healthcare, finance, and government where data retention restrictions are non-negotiable.

The research demonstrates that privacy-preserving AI isn't inherently inferior to unrestricted alternatives. As source code releases become standard practice, the methodology could accelerate adoption of responsible AI practices across industries. Future developments may focus on extending source-free unlearning to other architectures and improving efficiency for large-scale deployments.

• →SPACE enables unlearning in MLLMs without access to original training data, solving a critical privacy compliance challenge
• →The framework theoretically guarantees bounded perturbations on retained knowledge while maximizing feature spectral entropy
• →Experimental results show performance comparable to data-dependent methods across six datasets
• →Source-free unlearning addresses regulatory constraints under GDPR and emerging AI privacy frameworks
• →The approach uses text-guided proxy anchors and dual-constraint semantic isolation for indirect concept erasure