EigenShield introduces a computationally efficient, architecture-independent defense mechanism that protects Vision-Language Models (VLMs) from adversarial attacks by filtering out malicious signals.

• Leverages Random Matrix Theory to identify and neutralize adversarial disruptions in model representations
• Works at inference-time without requiring expensive retraining or model modifications
• Demonstrates superior performance compared to existing defenses against adaptive attacks
• Provides a theoretical foundation for understanding adversarial vulnerabilities in multimodal models

This research significantly advances AI security by addressing critical vulnerabilities in increasingly deployed multimodal systems, offering a practical defense that can be implemented across various model architectures.

Original Paper: EigenShield: Causal Subspace Filtering via Random Matrix Theory for Adversarially Robust Vision-Language Models