Xin Wang

The widespread practice of fine-tuning open-source Vision-Language Models (VLMs) raises a critical security concern: jailbreak vulnerabilities in base models may persist in downstream variants, enabling transferable attacks across fine-tuned systems. To investigate this risk, we propose the Simulated Ensemble Attack (SEA), a grey-box jailbreak framework that assumes full access to the base VLM but no knowledge of the fine-tuned target. SEA enhances transferability via Fine-tuning Trajectory Simulation (FTS), which models bounded parameter variations in the vision encoder, and Targeted Prompt Guidance (TPG), which stabilizes adversarial optimization through auxiliary textual guidance. Experiments on the Qwen2-VL family demonstrate that SEA achieves consistently high transfer success and toxicity rates across diverse fine-tuned variants, including safety-enhanced models, while standard PGD-based image jailbreaks exhibit negligible transferability. Further analysis reveals that fine-tuning primarily induces localized parameter shifts around the base model, explaining why attacks optimized over a simulated neighborhood transfer effectively. We also show that SEA generalizes across different base generations (e.g., Qwen2.5/3-VL), indicating that its effectiveness arises from shared fine-tuning-induced behaviors rather than architecture- or initialization-specific factors.

Understanding the memorization and privacy leakage risks in Contrastive Language--Image Pretraining (CLIP) is critical for ensuring the security of multimodal models. Recent studies have demonstrated the feasibility of extracting sensitive training examples from diffusion models, with conditional diffusion models exhibiting a stronger tendency to memorize and leak information. In this work, we investigate data memorization and extraction risks in CLIP through the lens of CLIP inversion, a process that aims to reconstruct training images from text prompts. To this end, we introduce \textbf{LeakyCLIP}, a novel attack framework designed to achieve high-quality, semantically accurate image reconstruction from CLIP embeddings. We identify three key challenges in CLIP inversion: 1) non-robust features, 2) limited visual semantics in text embeddings, and 3) low reconstruction fidelity. To address these challenges, LeakyCLIP employs 1) adversarial fine-tuning to enhance optimization smoothness, 2) linear transformation-based embedding alignment, and 3) Stable Diffusion-based refinement to improve fidelity. Empirical results demonstrate the superiority of LeakyCLIP, achieving over 258% improvement in Structural Similarity Index Measure (SSIM) for ViT-B-16 compared to baseline methods on LAION-2B subset. Furthermore, we uncover a pervasive leakage risk, showing that training data membership can even be successfully inferred from the metrics of low-fidelity reconstructions. Our work introduces a practical method for CLIP inversion while offering novel insights into the nature and scope of privacy risks in multimodal models.