arXiv Computer Vision

No Safe Dose: How Training Data Drives Unsafe Image Generation

The study investigates how the proportion of unsafe images in training data influences the safety of text‑to‑image models. By training identical models on corpora ranging from 0% to 9.6% unsafe content and evaluating 10,000 generations per model, the authors find that unsafe output increases monotonically with contamination, saturating at 26.4% for the highest contamination level. Filtering a fixed fraction of unsafe images is equally effective across different dataset sizes and does not degrade image quality, yet it cannot eliminate unsafe generation entirely, indicating that data curation and encoder choices are complementary but insufficient alone.

arXiv AI
Sep 24

InGuard: Towards Generalized Inner Guardrail for Safe Text-to-Image Generation

InGuard introduces an inner guardrail for text-to-image generation that operates within the model’s own representations, avoiding external classifiers. It grades prompts using the text encoder’s embeddings, modifies risky embeddings with SAGE to produce safe images, and employs a latent detector to halt generation early. Evaluated on the RevGen Safety Benchmark, InGuard achieves a 97.9–98.8% safety rate across five open-weight models while reducing benign disturbances, model parameters, and denoising steps.

By Zeyu Wang, Xiaodan Li, Zhiwen Li, Yuefeng Chen, Hui Xue
arXiv Computer Vision
Oct 16

Decoupled Safety Control: A Safety-Control Algorithm for Training-Free Safety Guidance

Decoupled Safety Control (DSC) is a plug‑in algorithm that improves training‑free safety guidance for diffusion models by separating the unsafe condition from the intervention process. Unlike existing coupled safety‑control chains, DSC determines the intervention based on the current generation state, reducing residual unsafe content and preserving benign‑generation quality. Experiments on standard safety benchmarks demonstrate that DSC achieves stronger unsafe‑content suppression while maintaining efficient inference.

By Huilin Zhou, Ruoxi Cheng, Yuhang Wang, Yuming Liu, Minghao Sun, Ruolong Ma, Lan Zhang
arXiv AI
Aug 19

DiSCO: Defending text-to-image generation through distribution-guided contrastive prompt optimization

DiSCO is a zero‑shot, black‑box defense for text‑to‑image models that operates solely at the prompt level. It expands prompts with a distribution‑guided suffix using beam search and contrastive scoring against safe and unsafe image pools generated by the target model, iteratively refining until safe content is produced. The method improves safety on the I2P benchmark under various red‑teaming attacks, reducing attack success rates by 37.7% and 25.13% while preserving semantic fidelity and image coherence.

By Tong Zhang, Motasem Alfarra, Carlos Hinojosa, Christos Louizos, Bernard Ghanem
arXiv AI
Jun 6

Do Models Share Safety Representations? Cross-Model Steering for Safe Visual Generation

arXiv:2606. 05290v1 Announce Type: cross Abstract: Recent progress in generative modeling has made safety control a central challenge, yet existing approaches remain largely model-specific, requiring retraining or tailored interventions for each new architecture.

By Tobia Poppi, Silvia Cappelletti, Sara Sarto, Florian Schiffers, Garin Kessler, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
arXiv AI
Oct 5

Corrupted but Correct: Why Vision-Language Models Lie to Themselves Internally

The paper demonstrates that a targeted adversarial perturbation can reduce a vision‑language model’s training loss to near zero for a fixed target caption, yet the same model, when generating freely, still produces the correct description. This phenomenon, termed the train/inference gap, is traced to a single autoregressive step where the target token’s rank is fixed across all images, and further analysis shows that the language decoder, rather than the visual encoder, determines whether the corrupted signal is amplified or suppressed. The study uses a controlled two‑stage PGD attack on Qwen2.5‑VL‑7B‑Instruct and evaluates the effect on 200 held‑out COCO images, revealing that adversarial robustness in autoregressive VLMs largely depends on the language decoder’s prior. whyItMatters":"The findings suggest that defenses and faithfulness evaluations for deployed vision‑language models should focus on the language decoder rather than the visual encoder, as the former is the key determinant of robustness to adversarial perturbations."

By Arun Josephraj Arokiaraj, Zekun Wu, Adriano Koshiyama
arXiv AI
Aug 25

A Reproducible, License-Aware Distillation Recipe for CPUDeployable Safety Classification

arXiv:2608.21570v1 Announce Type: new Abstract: Deploying a safety layer for large language models on commodity hardware is constrained by the guards available to do it: current open guard models hol...

By Edson Rodrigues da Cruz Filho, Paulo Ricardo Ferreira Neves, Paulo Henrique Eleuterio Falsetti, Jo\~ao Vitor Pavan, Ian Degaspari, Henrique Vieira Laturrague, Patrick Vieira Laturrague, Guilherme Nielsen Dias, Marccello Wilson Perez Berto, Gustavo Voltani Von Atzingen
arXiv Computation and Language
Oct 15

DecisionRobust: Benchmarking Robustness for JEV Style Decision Models Across Text and Vision

DecisionRobust is a new benchmark that evaluates the robustness of JEV-style decision models to adversarial attacks across both text and vision inputs. It includes a variety of attack types—semantic text perturbations, image perturbations, and prompt injections—and uses external validity checks and cross-model candidate selection to curate evaluation sets. Experiments on 30 text and 17 vision model configurations reveal that vulnerabilities can transfer across models, robustness varies by task and attack, and label-consistency checks significantly reduce apparent attack success.

By Yifan Lan, Zhongye Liu, Yurui Chang, Hanyu Wang, Yongkang Du, Lu Lin, Jinghui Chen
Hugging Face Trending Papers
Jul 7

AEGIS: A Mechanism-Guided Defense against Visual Synonym Jailbreaks in Text-to-Image Models

Text-to-image diffusion models have achieved high visual fidelity and broad adoption, but remain vulnerable to safety violations when adversaries exploit them to synthesize illicit content. Existing alignment paradigms, from input sanitization to structural feature pruning, are largely organized around unsafe concepts explicitly exposed during filtering, editing, or localization.