FORCE: Transferable Visual Jailbreaking Attacks via Feature Over-Reliance CorrEction

Abstract: The integration of new modalities enhances the capabilities of multimodal LLMs (MLLMs) but also introduces additional vulnerabilities. In particular, simple visual jailbreaking attacks can manipulate open-source MLLMs more readily than sophisticated textual attacks. However, these underdeveloped attacks exhibit extremely limited cross-model transferability, failing to reliably identify vulnerabilities in closed-source MLLMs. In this work, we analyse the loss landscape of these jailbreaking attacks and find that the generated attacks tend to reside in high-sharpness regions, whose effectiveness is highly sensitive to even minor parameter changes during transfer. To further explain the high-sharpness localisations, we analyse their feature representations in both the intermediate layers and the spectral domain, revealing an improper reliance on narrow layer representations and semantically poor frequency components. Building on this, we propose a Feature Over-Reliance CorrEction (FORCE) method, which guides the attack to explore broader feasible regions across layer features and rescales the influence of frequency features according to their semantic content. By eliminating non-generalizable reliance on both layer and spectral features, our method discovers flattened feasible regions for visual jailbreaking attacks, thereby improving cross-model transferability. Extensive experiments demonstrate that our approach effectively facilitates visual red-teaming evaluations against closed-source MLLMs.

• The paper introduces FORCE, a novel method that reduces model feature over-reliance to improve attack transferability.
• It combines layer feature regularization with spectral rescaling, achieving up to an 18.9% improvement in attack success rate and reducing query counts.
• The study lays a foundation for robust red-teaming in multimodal language models and highlights the need for improved defensive strategies.

Overview of "FORCE: Transferable Visual Jailbreaking Attacks via Feature Over-Reliance CorrEction"

This paper introduces a novel method named Feature Over-Reliance CorrEction (FORCE), which effectively enhances the transferability of visual jailbreaking attacks on multimodal LLMs (MLLMs). The method addresses the poor generalization of these attacks by correcting non-generalizable dependencies on model-specific features. The authors have conducted extensive experiments to demonstrate the enhanced transferability of attacks using FORCE compared to existing methods.

Motivation

The increasing capabilities of MLLMs through multimodal integration also open them up to vulnerabilities such as jailbreaking attacks. These attacks can exploit gaps in MLLM defenses, but often suffer from poor transferability across models. Previous attacks reside in high-sharpness regions in the input space, meaning their effectiveness diminishes significantly with even slight changes in model parameters or inputs. The core idea of this paper is to flatten these hypersensitivity regions and reduce over-reliance on specific features.

Methodology

The FORCE method involves two main strategies:

1. Layer Feature Regularization:
   • This component focuses on reducing dependence on narrow layer features. Through a regularization strategy applied across model layers, it encourages the attack to span broader regions of the feature space, promoting smoother loss landscapes.
2. Spectral Rescaling:
   • It addresses the reliance on high-frequency elements in the spectral domain that generally possess less semantic meaning. By guiding the attack to realign with natural frequency distributions, the method enhances feature generalizability.

     Figure 1: Schematic illustration of the generation and transfer of optimisation-based visual jailbreaking attacks, as well as the feasible regions of such attacks in the input space.

Experimental Evaluation

The experiments reveal substantial improvements in the transferability of visual attacks:

• Using FORCE, the attack success rate (ASR) saw consistent improvements across various target MLLMs, notably enhancing performance even in challenging scenarios involving early-fusion and commercial models.
• Across diverse datasets and architectures, FORCE demonstrated an average ASR improvement of 18.9% in some configurations, alongside significant reductions in the number of queries needed for a successful attack.

Figure 2: Feasible regions between jailbreaking and natural examples across different layers’ features. The blue and yellow points correspond to successful and failed examples on the source MLLM.

Ablation Studies

The ablation studies highlighted the individual contributions of each component of FORCE. Both layer feature regularization and spectral rescaling independently improved transferability, but their combined effect achieved the maximum performance gains.

Implications and Future Work

FORCE paves the way for more reliable and transferable visual jailbreaking attacks, offering a robust framework for red-teaming evaluations of MLLMs. However, the capabilities of FORCE highlight the need for improved defense strategies in multimodal systems. Future work may explore further integration with defense mechanisms to address the vulnerabilities exposed by such advanced attack methods. Additionally, expanding the scope of FORCE to other forms of adversarial attacks could present new insights into the defense and robustness of AI models.

Conclusion

FORCE represents a significant advancement in the field of adversarial machine learning, particularly for multimodal AI systems. By systematically addressing the limitations of current visual jailbreaking attacks, it enhances their real-world applicability in evaluating the security of complex AI systems. The study underscores the importance of understanding internal feature dependencies to improve cross-model attack transferability.