Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Multimodal Generative Engine Optimization: Rank Manipulation for Vision-Language Model Rankers
TL;DR
Vision-Language Models (VLMs) struggle with reliably ranking multimodal items due to vulnerabilities in their cross-modal knowledge application. The authors developed Multimodal Generative Engine Optimization (MGEO), a technique that manipulates VLM rankings by crafting subtle image changes and coherent text.
✦ Why It Matters
Engineers should consider the vulnerabilities of VLMs when designing retrieval systems and explore defense strategies.
Key Takeaways
- Investigate vulnerabilities in Vision-Language Models for ranking tasks.
- Implement Multimodal Generative Engine Optimization techniques in testing.
- Focus on internal knowledge mechanisms for effective model design.
- Develop defense strategies against adversarial manipulations in multimodal systems.
Full Summary
Vision-Language Models (VLMs) combine visual and textual information to enhance retrieval and recommendation systems, but their ability to rank multimodal items accurately is questionable. The study introduces Multimodal Generative Engine Optimization (MGEO), which allows adversaries to manipulate VLM rankings by creating imperceptible image perturbations alongside fluent textual additions.
This approach employs an alternating optimization strategy to exploit the deep interactions between visual and linguistic representations within the VLM. Results showed that MGEO achieved rank manipulations that were significantly more effective than those from unimodal attacks and heuristic baselines, indicating that surface-level content quality alone is insufficient for rank improvement.
The findings highlight the need for a deeper understanding of how VLMs utilize their internal knowledge, raising concerns about the robustness of these models. This research encourages the development of defense mechanisms to protect multimodal retrieval systems from such vulnerabilities.