Cite as: ArXiv(2505.15130) Submitted on 2025/09 Abstract 問題：CLIP 這類大型 VLM 模型很容易受到 Adversarial Attack 攻擊 。 雖然 Full Fine-Tuning 可以提升防禦力，但成本太高；而現有的 PEFT 方法（如 Adversarial Prompt Tuning）在 Few-shot 的...

Cite as: ACL 2024 Submitted on 2023/11 Problems Confounder Memes（混淆梗圖）問題: 圖像或文本的微小差異可能導致完全不同的含義 現有CLIP-based系統的embedding space對細微差異不敏感 混淆梗圖在embedding space中距離過近，難以區分 → 論文發現HateCLIPper等現有模型將相反...

Papers CapAlign: Improving Cross Modal Alignment via Informative Captioning for Harmful Meme Detection Cite as ArXiv:2411.07527 Submitted on 2025/05 REF: https://a...

Cite as ArXiv:2411.07527 Submitted on 2025/05 REF: https://arxiv.org/abs/2411.07527 Introduction (Hateful meme classification) Stage 1: 早期的多模態融合 (Alignment and Fusion) 迷因包含「圖片」和「文字」兩種模態，所以模型必須...

1. (DiffPure) Diffusion Models for Adversarial Purification ICML (2022) Submitted on 2022/04 2. (GrIDPure) Can Protective Perturbation Safeguard Personal Data from Being Exploited by Stable Diffu...

SaRA (High-Efficient Diffusion Model Fine-tuning with Progressive Sparse Low-Rank Adaptation) Cite as: ICLR 2025 Submitted on 2025/04 Abstract SaRA是一種用於預訓練擴散模型的高效微調方法。 SaRA旨在重新利用這些無效參數，使其重新變得有...

IMPRESS: Evaluating the Resilience of Imperceptible Perturbations Against Unauthorized Data Usage in Diffusion-Based Generative AI NeurIPS (2023) Submitted on 2023/10 CA...

ArXiv: 2503.05595 Submitted on 2025/03 REF: https://arxiv.org/abs/2503.05595 Abstract Extended Defense Scope: Expand the defense to include both tuning-based and editing-based methods, while ot...

Cite as: NeurIPS 2024 REF: https://arxiv.org/pdf/2503.09124 Abstract Novel Approach: Models adversarial attacks as a non-parametric diffusion process instead of using traditional denoisi...

Cite as: arXiv Submitted on 2025/03 REF: https://arxiv.org/abs/2503.13945 Comparison across different black-box prompts Anti-Dreambooth (Anti-DB) Alternating Surrogate and Perturbation Learnin...