In this paper, we show for the first time that visual token pruning enhances the robustness of Multimodal Large Language Models (MLLMs), mitigating vulnerabilities such as jailbreak attacks and hallucinations. Given that vision and language modalities cannot be perfectly aligned, the misaligned visual tokens might act as out-of-distribution (OOD) inputs, leading to unpredictable outputs and introducing potential vulnerabilities. Building on this insight, we aim to enhance model robustness against jailbreaks and hallucinations by reducing OOD visual tokens at robust-pruning layers, while also reducing inference cost as a side benefit. Specifically, we measure the distance between each visual token and the language feature space. Then, visual tokens with large distances are identified as OOD tokens, which can be iteratively pruned. To demonstrate the effectiveness of our method, we evaluate it on seven diverse popular benchmarks. Notably, our method yields an average improvement of 13.29\% in defending jailbreak attacks, consistently achieves competitive performance in mitigating hallucinations, and maintains strong results on general datasets like MME.
Shishen Gu, Jie-Quan Cui, Wenbo Hu et al.· arXiv.org· 1 citation
Existing semi-supervised learning (SSL) methods rely predominantly on pseudo-labeling and consistency regularization to leverage unlabeled data, demonstrating significant performance improvements. However, we pinpoint that these methods suffer from a confidence-for-weighting issue, overvaluing high-confidence pseudo-labels while undervaluing low-confidence yet informative samples that are critical for robust generalization. In this paper, we introduce EntropyMatch, an entropy-driven SSL framework that redefines sample importance through prediction entropy rather than confidence alone. EntropyMatch employs a bidirectional weighting strategy: upward exploitation exploits reliable hard samples to refine decision boundaries while downward exploration cautiously explores uncertain ones to reduce noise. Additionally, EntropyMatch features an adaptive training mechanism that aligns with model maturity, shifting focus from safe exploration to strategic exploitation as training progresses. Experiments on eight benchmarks across various SSL tasks—spanning image classification, facial expression recognition, and human action recognition—validate EntropyMatch’s robustness and effectiveness. It consistently achieves state-of-the-art results, notably matching state-of-the-art LION’s performance on RAF-DB with just half the labeled data, demonstrating superior data efficiency and generalization.
Anyang Tong, Zenglin Shi, Zhun Zhong et al.· IEEE Transactions on Image P...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
