Li Shen

Linghui Shen, Mingyue Cui, Xingyi Yang

Visual world models (VWMs) synthesize interactive, action-conditioned rollouts from a single context image. However, it remains an open question how robust these models are to adversarial perturbations. Standard adversarial attacks fail to assess this vulnerability because attackers lack ground-truth future videos and cannot predict subsequent user controls. We introduce BadWorld, a label-free adversarial framework tailored for autoregressive VWMs that systematically overcomes both constraints. First, to bypass the need for future supervision, we propose a self-supervised velocity attack that directly disrupts the early denoising dynamics of the model. Second, to ensure the attack generalizes across unpredictable user actions, we formulate a trajectory-adaptive bi-level optimization that actively mines hard control sequences to forge control-agnostic perturbations. Evaluated on representative VWMs with continuous and discrete controls, BadWorld exposes severe structural fragility. Visually indistinguishable adversarial images reliably trigger catastrophic degradation in future rollouts, leading to incomplete denoising, structural collapse, and control inconsistency. These findings reveal critical risks for deploying VWMs in safety-critical systems while highlighting a practical mechanism for privacy protection.

Mingyue Cui, Linghui Shen, Xingyi Yang

Sparse Autoencoders (SAEs) decompose residual-stream activations into interpretable features. Recent latent-space defenses increasingly rely on these decompositions, assuming that identified "unsafe" SAE features serve as actionable handles for monitoring and intervention. In this paradigm, clamping a specific harmful feature is expected to reliably prevent model misbehavior. However, we show that this success may hide a recoverable failure mode: the clamp may block one visible route to a behavior without eliminating the behavior itself. We formulate this vulnerability as post-intervention recovery, a constrained residual-space optimization problem. Starting from the post-intervention residual state, we optimize residual perturbations to recover the pre-intervention behavior while preserving the post-intervention values of the targeted SAE features. Even under a strong threat model where the intervention remains active throughout optimization and generation, recovery remains possible. To rule out that recovery simply undoes the intervention, we use encoder-orthogonal updates for single-layer interventions and the corresponding feature-map Jacobian in the cross-layer setting. Across TPP, unlearning, IOI, and refusal steering experiments, this stress test reveals recoverable behavior despite successful feature-level intervention. Especially in the safety-critical refusal-steering setting, we achieve a 95.8% recovery rate on valid samples while keeping defended-feature relative drift to 0.131, substantially below suffix-based baselines. A recovery-path attribution analysis further localizes this recovery to the SAE reconstruction residual, the component left unexplained by the SAE. These results expose a gap between feature-level control and behavioral completeness: SAE features can support causal intervention, but controlling them does not guarantee control over the underlying behavior.

Yuni Lai, Xiaoyu Xue, Linghui Shen et al.

Language-empowered foundation models (LeFMs), such as CLIP and GraphCLIP, have transformed multimodal learning by aligning visual (or graph) features with textual representations, enabling powerful downstream capabilities like few-shot learning. However, the reliance on small, task-specific support datasets collected in open environments exposes these models to poisoning attacks, where adversaries manipulate the support samples to degrade performance. Existing defenses rely on empirical strategies, which lack formal guarantees and remain vulnerable to unseen and adaptive attacks. Certified robustness offers provable guarantees but has been largely unexplored for few-shot classifiers based on LeFMs. This study seeks to fill these critical gaps by proposing the first provably robust few-shot classifier that is tailored for LeFMs. We term our model Language-empowered Few-shot Certification (\textbf{LeFCert}). It integrates both textual and feature embeddings with an adaptive blending mechanism. To achieve provable robustness, we propose a twofold trimmed mean prototype and derive provable upper and lower bounds for classification scores, enabling certification under worst-case poisoning scenarios. To further enhance the performance, we extend LeFCert with two variants by considering a more realistic and tighter attack budget: LeFCert-L incorporates randomized smoothing to provide Lipschitz continuity and derive robustness under dual budget constraints, and LeFCert-C provides collective certification for scenarios where attackers distribute a shared poisoning budget across multiple samples. Experiments demonstrate that LeFCert achieves state-of-the-art performance, significantly improving both clean and certified accuracy compared to existing baselines. Despite its advanced robustness mechanisms, LeFCert is computationally efficient, making it practical for real-world applications.