SafeIMG Logo  SafeIMG

**AI-generated Images Challenge Visual Trust in High-risk Scenarios** **Yi-Zhi Wang1,2, Yichen Xiao1,2, Linan Yue1,2, Weibo Gao3, Yichao Du4, Pengfei Fang1,2, Shimin Di1,2, Min-Ling Zhang1,2** 1 Southeast University    2 Key Laboratory of Computer Network and Information Integration, Ministry of Education 3 The Hong Kong Polytechnic University    4 School of Artificial Intelligence, Wuhan University [![Project Page](https://img.shields.io/badge/Project-Page-0b6b45)](https://safeimg.github.io/) [![Dataset](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-f2b134)](https://huggingface.co/datasets/Snowstorm1492/SafeIMG) [![Task](https://img.shields.io/badge/Task-AI--generated%20Image%20Detection-7a5af8)](#benchmark-tasks) [Overview](#-overview) Β· [Dataset](#-dataset-examples) Β· [Results](#-experimental-results) Β· [Quick Start](#-quick-start) Β· [Citation](#-citation) **SafeIMG is a safety-oriented benchmark for detecting AI-generated images that imitate visual evidence in high-risk public- and individual-safety scenarios.**
> [!WARNING] > SafeIMG contains synthetic depictions of disasters, accidents, violence, emergencies, transactions, and other sensitive scenarios. The data are intended for research and evaluation only. Images in this benchmark must not be presented as records of real events. # πŸ“‘ News - **2026-07:** SafeIMG dataset and evaluation code released. - **2026-07:** SafeIMG paper released. ## πŸ“‹ Overview Modern image generators can create realistic, story-rich images that resemble news photographs, transaction records, chat screenshots, identity endorsements, and other forms of visual evidence. Existing synthetic-image detection benchmarks primarily cover general natural images, earlier generators, or domains that are not directly tied to safety-critical decisions. SafeIMG asks a more demanding question: > **Can a detector recognize AI-generated images when they imitate evidence used in high-stakes real-world decisions?** SafeIMG provides: - **12 high-risk scenarios** spanning public safety and individual safety; - images generated with **GPT Image 2**, representing a challenging new-generator distribution; - a risk-oriented scenario taxonomy and structured, story-complete prompts; - fine-grained human annotations of suspicious regions, artifact types, and textual explanations; - evaluation of specialized forensic detectors, multimodal large language models, and humans; - diagnostic analyses of explanation alignment and robustness under network-propagation degradation.

Overview and representative samples of SafeIMG

## πŸ’‘ Why SafeIMG? 1. **Safety-critical visual evidence.** SafeIMG focuses on images whose misuse may affect public judgment, emergency response, financial trust, identity claims, personal reputation, or legal interpretation. 2. **A modern generator distribution.** Samples are produced with GPT Image 2 rather than only GANs or early diffusion models, exposing substantial distribution shift for existing detectors. 3. **Diagnosis beyond a real/fake label.** Each annotated artifact records where the suspicious evidence is, what type of issue it represents, and why it appears abnormal. 4. **Local and high-level reasoning.** The benchmark covers explicit artifacts involving text, faces, hands, or lighting as well as scene-level commonsense conflicts and violations of physical laws. 5. **Realistic dissemination conditions.** SafeIMG measures robustness after compression and image degradation that resemble online sharing and re-encoding. # πŸ“š Benchmark Tasks SafeIMG supports four complementary evaluation tracks: | Track | Input | Expected output | Main purpose | |---|---|---|---| | Image authenticity | One image | `real` or `ai_generated` | Measure image-level detection capability | | Artifact diagnosis | One generated image | Suspicious regions and artifact types | Test whether a detector identifies meaningful forensic cues | | Rationale alignment | Image plus model explanations | Atomic visual reasons | Compare model evidence with human annotations | | Propagation robustness | Degraded or re-encoded image | Authenticity prediction | Test stability under realistic dissemination transformations | ## 🧩 Taxonomy SafeIMG contains eight public-safety categories and four individual-safety categories. | ID | Domain | Category | Evidentiary function | |---|---|---|---| | P1 | Public safety | Natural Disasters | Claims about disaster impact and emergency response | | P2 | Public safety | Social Unrest, Public Violence, and Attack Incidents | Claims about public order, security risks, and social stability | | P3 | Public safety | Transportation Accidents | Claims about occurrence, damage, and responsibility | | P4 | Public safety | Fires, Explosions, and Energy Facility Accidents | Risk assessment and emergency handling | | P5 | Public safety | Pollution and Hazardous Material Accidents | Claims that may trigger panic or regulatory action | | P6 | Public safety | Building and Civil Infrastructure Accidents | Claims about infrastructure safety, damage, and liability | | P7 | Public safety | Crowd Gathering and Venue Safety Accidents | Crowd control, evacuation, and event-safety judgments | | P8 | Public safety | Public Health and Biosecurity Events | Claims affecting public-health decisions and social trust | | I1 | Individual safety | Personal Accidents and Emergencies | Claims about personal danger, rescue needs, or liability | | I2 | Individual safety | Private Receipts and Transaction Records | Claims about payments, refunds, reimbursement, or disputes | | I3 | Individual safety | Personal Chat and Communication Records | Claims about private commitments, misconduct, or disputes | | I4 | Individual safety | Fabricated Scene Evidence and Identity Endorsement | Claims about presence, status, identity, or endorsement | ## πŸ› οΈ Benchmark Construction SafeIMG is built in five stages: 1. **Risk-oriented taxonomy:** define public- and individual-safety categories. 2. **Scenario-space specification:** specify a risk summary, content list, and media-form list for each category. 3. **Prompt construction:** expand each scenario with a 5W+1H narrative structure and category-specific constraints. 4. **Image generation and filtering:** generate images with GPT Image 2 and remove invalid or severely off-topic samples. 5. **Human annotation:** localize suspicious regions, assign artifact labels, and write concise explanations.

Five-stage construction pipeline of SafeIMG

## πŸ“ Annotation Design For each retained synthetic image, SafeIMG preserves construction metadata such as its risk category, scenario instance, complete generation prompt, and media form. Fine-grained annotations may additionally contain: - one or more suspicious regions; - an artifact type for each region, such as `Commonsense`, `Faces`, `Fonts`, `Physics`, `Hands`, `Lighting`, or `Other`; - a natural-language explanation describing the concrete issue; - an image-level authenticity label. This design enables evaluation at both the **decision level** (is the image real or generated?) and the **evidence level** (where is the problem, and why is it suspicious?). ## 🌟 Dataset Examples SafeIMG includes documentary-style images, news-like photographs, transaction records, device screenshots, chat records, and identity-endorsement material. The overview above shows representative samples from all 12 categories together with region-level artifact annotations. The following snippet loads one record and inspects the exact schema published on Hugging Face: ```python from datasets import load_dataset dataset = load_dataset( "Snowstorm1492/SafeIMG", split="test" ) print(dataset) print(len(dataset)) ``` ## πŸ“Š Experimental Results SafeIMG exposes a substantial gap between current automated detectors and human forensic judgment.

Five-stage construction pipeline of SafeIMG

Five-stage construction pipeline of SafeIMG

## πŸš€ Quick Start ### 1. Set up Clone the repository and enter the project directory: ```bash git clone https://github.com/Snowstorm1492/SafeIMG.git cd SafeIMG ``` We recommend creating a new Conda environment: ```bash conda create -n safeimg python=3.11 conda activate safeimg ``` Install the required dependencies: ```bash pip install -r requirements.txt ``` ### 2. Load SafeIMG from Hugging Face ```python from datasets import load_dataset dataset = load_dataset( "Snowstorm1492/SafeIMG", split="test" ) print(dataset) print(len(dataset)) ``` To download a local snapshot instead: ```bash hf download Snowstorm1492/SafeIMG \ --repo-type dataset \ --local-dir data/SafeIMG ``` ### 3. Base VLM evaluation Run the base VLM evaluation with the following command: ```bash python evaluation_vlm.py \ --base-url "" \ --api-key "" \ --model "" \ --output "results/predictions.jsonl" ``` ### 4. Specialized detector evaluation Run the specialized AI detection model evaluation with the following command. CNNSpot, FreDect, and LNP are supported. ```bash python evaluation_spec.py \ --methods cnnspot fredect lnp \ --dataset Snowstorm1492/SafeIMG \ --weights-dir weights \ --device cuda \ --lnp-device cpu \ --output-dir results/specialized ``` The required weights and auxiliary files can be downloaded from the official repositories of CNNSpot, FreDect, LNP or [AIGCDetectBenchmark](https://github.com/Ekko-zn/AIGCDetectBenchmark). The expected directory structure is as follows: ```text weights/ β”œβ”€β”€ classifier/ β”‚ β”œβ”€β”€ CNNSpot.pth β”‚ β”œβ”€β”€ FreDect.pth β”‚ └── LNP.pth β”œβ”€β”€ preprocessing/ β”‚ └── sidd_rgb.pth └── auxiliary/ β”œβ”€β”€ dct_mean └── dct_var ``` ### 5. Human - AI alignment in artifact explanations Evaluate human–AI alignment in artifact explanations with the following commands. Evaluate `human_covered_by_ai`: ```bash python evaluation_human_covered_by_ai.py \ --predictions "results/predictions.jsonl" \ --base-url "" \ --api-key "" \ --model "gpt-5.5" ``` Evaluate `ai_supported_by_human`: ```bash python evaluation_ai_supported_by_human.py \ --predictions "results/predictions.jsonl" \ --base-url "" \ --api-key "" \ --model "gpt-5.5" ``` ## βš™οΈ Responsible Use SafeIMG is designed to improve research on visual authenticity and forensic reliability. It must not be used to fabricate evidence, impersonate individuals, misrepresent synthetic events as real, or facilitate fraud, harassment, or panic. Benchmark performance should not be interpreted as proof that a model is suitable for autonomous deployment in legal, financial, medical, emergency-response, or public-safety decisions. Current systems can miss realistic synthetic images and can also falsely accuse real images of being generated. Human review, provenance information, source verification, and contextual evidence remain necessary. ## πŸ™ Acknowledgments We thank the open-source communities behind [Hugging Face](https://huggingface.co/) and [PyTorch](https://pytorch.org/) for their valuable tools and infrastructure. We also thank the authors and contributors of [CNNSpot](https://github.com/PeterWang512/CNNDetection), [FreDect](https://github.com/RUB-SysSec/GANDCTAnalysis), [LNP](https://github.com/Tangsenghenshou/Detecting-Generated-Images-by-Real-Images), and [AIGCDetectBenchmark](https://github.com/Ekko-zn/AIGCDetectBenchmark), whose publicly available implementations informed our specialized detector evaluation. ## πŸ›‘οΈ License SafeIMG is licensed under the [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License](https://creativecommons.org/licenses/by-nc-sa/4.0/). ## πŸ”— Citation If SafeIMG is useful in your research, please cite: ```bibtex @misc{wang2026aigeneratedimageschallengevisual, title={AI-generated Images Challenge Visual Trust in High-risk Scenarios}, author={Yi-Zhi Wang and Yichen Xiao and Linan Yue and Weibo Gao and Yichao Du and Pengfei Fang and Shimin Di and Min-Ling Zhang}, year={2026}, eprint={2607.22745}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2607.22745}, } ``` For questions about the benchmark, data, or evaluation protocol, please contact **Linan Yue** at [lnyue@seu.edu.cn](mailto:lnyue@seu.edu.cn) or open a GitHub issue.