This research advances security in AI systems by improving the detection of anomalous inputs through cross-modal representation alignment.

• Addresses limitations in existing multi-modal out-of-distribution detection methods
• Optimizes pretrained vision-language models beyond frozen weights
• Enhances detection of potential security threats and adversarial examples
• Provides superior performance for downstream security applications

This approach significantly strengthens AI security frameworks by better identifying inputs that differ from training data, helping prevent unauthorized access and improving system robustness against potential attacks.

Enhanced OoD Detection through Cross-Modal Alignment of Multi-Modal Representations