TL;DR
This paper extends Randomized Smoothing (RS), a powerful technique for certifying the robustness of classification models, to Vision-Language Models (VLMs). By reformulating VLM outputs as discrete classes via an "oracle classifier," they enable provable robustness guarantees against image perturbations for VLMs, a significant step towards safer generative AI.
What Problem Does This Paper Address?
Deep learning models, especially large ones like Vision-Language Models (VLMs), are susceptible to adversarial attacks where small, imperceptible changes to the input can drastically alter the output. While techniques like Randomized Smoothing (RS) offer provable robustness guarantees for classification tasks, they haven't been directly applicable to generative models like VLMs because their outputs are sequences, not single labels. This gap leaves VLMs vulnerable to attacks, hindering their deployment in safety-critical applications. The paper aims to bridge this gap by making RS applicable to VLMs, allowing for formal verification of their robustness.
Methods
The core innovation is adapting RS to generative models by introducing an oracle classification layer. Instead of directly classifying the VLM's generated text, the VLM's output is passed to an auxiliary "oracle" model (e.g., a strong LLM) that classifies it into a discrete category. This allows RS to be applied. The paper proposes three main ways to use this oracle:
- Content Moderation: The oracle classifies VLM outputs as "harmful" or "harmless."
- VLAs with Discrete Actions: The oracle maps VLM outputs to predefined actions (e.g., robot commands).
- Semantically Equivalent Output Clustering: The oracle groups semantically similar VLM outputs into equivalence classes.
For these approaches, the paper develops theoretical frameworks to derive certified radii and analyzes scaling laws for sample efficiency. They show that their method can significantly reduce the number of samples needed for certification compared to previous RS work, making it more practical for large VLMs.
Datasets
The paper does not introduce new datasets. Instead, they evaluate their framework on state-of-the-art VLMs using prompts and adversarial attacks from existing benchmarks, specifically referencing the "harmful prompts" and attack methodology from Qi et al. (2024). They use LLaVA 1.6 as the VLM and Gemma 2 (9b version) as the oracle LLM.
Key Insights and Ideas
- Bridging the Generative-Robustness Gap: The most significant insight is that by framing VLM outputs through an oracle classifier, we can leverage the established theory of RS for classification. This is a crucial conceptual leap.
- Sample Efficiency is Key: The paper demonstrates that their approach significantly reduces the sample complexity required for certification, making it computationally feasible for large VLMs. They show that 2-3 orders of magnitude fewer samples can suffice with minimal loss in robustness.
- Oracle Imperfection is Manageable: The theoretical analysis accounts for the oracle classifier not being perfect, providing bounds on the certified radius even when the oracle makes mistakes (as long as its error rate is below 0.5).
- Parallelization for Speed: The image perturbation approach allows for parallel processing of noisy samples, enabling efficient batched inference and reducing overall certification time.
Key Results
- The proposed framework is experimentally validated on state-of-the-art VLMs, demonstrating certified robustness against recent jailbreak-style adversarial attacks.
- The theoretical predictions for certified radius reduction and accuracy drop align well with empirical results, confirming the effectiveness of their scaling laws.
- They show that using around 100 samples can achieve a significant portion (e.g., 60%) of the certified radius obtained with thousands of samples.
- Batched RS certification is shown to be efficient, with inference times remaining relatively constant for up to ~50 samples and increasing predictably for more samples.
Limitations and Future Work
- Textual Perturbations: The current work focuses on image perturbations. Extending RS to textual perturbations in VLMs is a significant open challenge, requiring new conceptual and algorithmic advances for handling "noisy prompts."
- Threat Models: The certified defenses are limited to threat models previously covered by RS (L2, Lp norms, geometric perturbations). Capturing every possible perturbation strategy remains a challenge for adversarial robustness literature in general.
- Oracle Accuracy: While the theory accounts for oracle imperfection, the practical performance still relies on a highly accurate oracle.
Future work is explicitly directed towards extending RS to text-based generative models and exploring more general threat models to make robustness certification more practically useful.
Summary updated