TL;DR
FusionAudit introduces a diagnostic-driven framework to uncover hidden vulnerabilities in modern "native" Vision-Language Models (VLMs) by targeting specific visual pathways. By moving beyond generic adversarial attacks to architecture-aware interventions like MANTIS and Nat-Evid, the authors demonstrate that standard benchmarks significantly underestimate the residual risk in deeply integrated multimodal systems.
What Problem Does This Paper Address?
The architecture of Vision-Language Models is shifting. Early "alignment-era" models (like LLaVA) were modular, essentially piping a frozen vision encoder into a language model via a simple projection layer. Modern "native" VLMs, however, feature deep fusion where visual and textual tokens interact in a unified representational space from the very first layers.
Current robustness evaluations haven't kept pace with this structural transition. They typically rely on "pixel-centric" attacks (like PGD) that treat the entire visual input as a single surface. This approach fails for native models because:
- Gradient Mismatch: Standard end-to-end attacks spread gradient pressure across all output tokens, often missing the specific "decision token" where a model commits to an error.
- Visual Rejection: Native models exhibit a "see-but-not-deceived" behavior—they might "read" an adversarial text overlay but "reject" it as a task response because it looks visually incongruent with the rest of the image.
Methods
The authors propose PaRMA (Pathway-conditioned Risk for Multimodal Auditing), a framework that decomposes the visual attack surface into two primary pathways: pixel perturbation ($\pi_{pix}$) and rendered-text insertion ($\pi_{txt}$).
To exploit the blind spots of native architectures, they introduce two novel attacks within their FusionAudit pipeline:
- MANTIS (Margin-Aware Native Token Intervention Strategy): Instead of optimizing against the full output sequence, MANTIS concentrates gradient pressure exclusively on the "first-divergence token"—the exact point where the correct and wrong answer sequences first differ.
- Nat-Evid (Naturalized Evidence): To bypass the model's visual rejection of simple typographic overlays, Nat-Evid uses generative image editing to insert "host-congruent" text (e.g., making a wrong answer look like a native chart annotation or scene text).
They also introduce two key diagnostics:
- Read–Use Gap (RUG): Measures the difference between a model's ability to read inserted text and its tendency to use it in a task response.
- Overlay Attention Ratio (OAR): Quantifies how much attention the model's decoder pays to an adversarial overlay relative to its size.
Datasets
The audit was conducted on a balanced 1500-sample VQA evaluation set compiled from three major benchmarks:
- ChartQA: Focuses on visual reasoning over structured data.
- DocVQA: Targets document understanding.
- MME: Provides a mix of natural image categories (artwork, celebrities) and text-dense reasoning.
The authors evaluated 18 open-weight VLMs categorized by a new "nativeness taxonomy" ranging from $L_1$ (modular) to $L_3$ (deep/early fusion).
Key Insights and Ideas
- The "See-But-Not-Deceived" Signature: In native VLMs, attention diagnostics (OAR) show that models often focus heavily on typographic overlays, and reading prompts prove they can "see" the text. However, they frequently ignore this text when answering the main task. This suggests native models have a fine-grained sensitivity to visual congruence that modular models lack.
- Objective Mismatch in Deep Fusion: In $L_3$ models, visual features re-enter the computation at multiple layers. Standard attacks anchored only at the vision encoder or spread across the whole output sequence are less effective than MANTIS, which targets the specific decision-making token.
- Ranking Instability: The paper reveals that legacy rankings based on standard PGD are unstable for native models. A model that appears robust under PGD might be highly vulnerable to MANTIS, meaning single-scalar benchmarks are no longer reliable for model selection.
Key Results
The diagnostic-driven attacks in FusionAudit significantly raised the estimated residual risk across all model types:
- Increased Risk Detection: Compared to baseline attacks, the diagnostic-driven suite increased the detected risk by +10.6% in $L_1$, +16.9% in $L_2$, and a staggering +25.1% in $L_3$ models.
- MANTIS Superiority: The fraction of examples broken exclusively by MANTIS (and missed by standard PGD) peaked at 24.6% for the most advanced $L_3$ models.
- Stability: Ablation studies showed that MANTIS is robust to hyperparameter choices like random initialization and optimization budget, confirming that its gains are structural rather than just "lucky" tuning.
Limitations and Future Work
- Observational Taxonomy: The $L_1$–$L_3$ levels are correlational; while they track fusion depth, other factors like training data and scale also vary between model families.
- The $L_4$ Gap: The authors define a theoretical $L_4$ level (strict single-backbone models like Chameleon), but these were not included in the empirical audit due to current limitations in handling text-heavy images.
- Scope of Attacks: The current framework focuses on correctness-based Attack Success Rate (ASR) for wrong-answer injection. It does not yet cover "jailbreaking" or safety-alignment bypasses, which require different, compliance-based metrics.
- Extensibility: The authors envision expanding PaRMA to other modalities like audio and video as native multimodal models continue to evolve.
Summary updated