Paper accepted at ACCV 2026
10 September 2026
We are excited to share our new paper, “On Adversarial Vulnerability of Vision-Language Models through the Lens of Intermediate Spectral Subspaces,” accepted at the 18th Asian Conference on Computer Vision (ACCV 2026).
This work is authored by Chethan Krishnamurthy Ramanaik, Tobias Callies, Michael Hecht, and Prof. Eirini Ntoutsi.
Vision-language models combine visual and textual information, but small, visually imperceptible changes to an image can still cause them to produce incorrect or unrelated responses. This paper investigates this vulnerability by examining the spectral structure of the models’ intermediate linear transformations.
Key takeaways:
-
A new perspective on adversarial vulnerability. The study identifies near-null, or bottom singular-vector, subspaces as a previously overlooked attack surface in vision-language models.
-
A new spectral attack. The proposed Spectral Subspace Guided Representation Attack (SSGRA) guides intermediate representations toward these information-attenuating directions.
-
Stronger attacks across different models. SSGRA outperformed six representative baselines on Qwen2.5-VL, LLaVA-1.5, and Gemma 3 while keeping the visual perturbations very small.
-
A connection between model structure and vulnerability. Models containing more near-null spectral directions were also found to be more susceptible to the proposed attack.
-
Implications for robustness. Controlling near-null singular directions could complement existing robustness techniques that focus mainly on limiting large singular values.
The findings provide both a stronger method for evaluating vision-language models and a new direction for designing more robust multimodal AI systems.