Research map / Detection and segmentation / Open-set and few-shot detection
MI-CAPTCHA: Enhance the Security of CAPTCHA Using Mooney Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HiMI, a framework that converts natural images into adjustable-difficulty Mooney images via CQNet color quantization and a perceived difficulty adjustment module, and uses them to build two CAPTCHA schemes (object detection and instance segmentation).
Problem
Image-based CAPTCHAs are increasingly broken by deep vision models; the authors ask whether Mooney images, which exploit gaps in human closed-loop perception, can strengthen CAPTCHA security.
Approach
A learned 1-bit color quantization network (CQNet, UNeXt autoencoder with VGG19 feature loss) plus a graph-based perceived difficulty adjustment over pixel-cluster adjacency graphs, with parameters alpha and beta controlling recognition cues and saliency, used to generate MIs for two CAPTCHA schemes.
Key results
- HiMI MIs achieve content loss 1.02 and style loss 0.19 versus 1.16 and 0.28 for the best baseline (AdaConv), and human recognition accuracy 0.937 in 13 s versus 0.359 at 35 s for AdaConv.
- On COCO, YOLO detection accuracy on HiMI MIs drops to 0.258 versus 0.412 for PhotoMI; for instance segmentation, YOLO mean IoU falls to 0.216 and SAM to 0.104.
- In Scheme 1 with M=2, human success rate is 0.763 while YOLO success rate is 0.105, the largest human-solver gap among settings tested.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HiMI-generated MIs instead of AdaConv, PhotoMI, or natural images | YOLO and DINO object detection · COCO and Animal 2K | accuracy: COCO YOLO: 0.673 -> 0.258; DINO: 0.669 -> 0.249; Animal 2K YOLO: 0.748 -> 0.324; DINO: 0.736 -> 0.311 | Helped |
| Use HiMI-generated MIs instead of natural images | YOLO and SAM instance segmentation · MI instance segmentation | mean IoU: YOLO: 0.974 -> 0.216; SAM: 0.986 -> 0.104 | Helped |
| Generate MIs with HiMI instead of baseline methods | MI generation · MI quality user study | accuracy: 0.325 -> 0.937 | Helped |
| Use CQNet instead of OcTree, MedianCut, or ColorCNN for color quantization (ablation) | HiMI · MI quality | content loss | Helped |
| Vary α and β in the PDA module to adjust perceived difficulty (ablation) | HiMI · MI quality | difference from natural images | Mixed |
| Set Scheme 1 parameter M from 1 to 4 (ablation) | MI-CAPTCHA Scheme 1 · MI-CAPTCHA Scheme 1 | success rate: Humans: 0.895 -> 0.379; YOLO: 0.313 -> 0 | Hurt |
| Set M=2 in Scheme 1 rather than M=1, 3, or 4 (ablation) | MI-CAPTCHA Scheme 1 · MI-CAPTCHA Scheme 1 | human success rate: 0.895; 0.763; 0.631; 0.379 | Mixed |
Surprising findings
- The UST baselines' binarized outputs were far less perceptible to humans (e.g., WCT accuracy 0.023), showing that generic style transfer does not yield usable MIs.
Limitations
- Human study uses only 120 participants and fixed sample sizes; security of schemes not evaluated against adversarial attacks or fine-tuned solvers beyond YOLO/DINO/SAM.
- Scheme 1 solver evaluation relies on fine-tuned detectors; dependence on segmentation model (GFM) quality for PDA preparation.
- The paper does not specify detailed CQNet training data or hyperparameters beyond loss layers.
Experimental setup
Datasets: Animal 2K MS COCO 2017
Benchmarks: MS COCO 2017 val Animal 2K test
Metrics: content loss style loss recognition accuracy recognition time mAP-style detection accuracy mean IoU success rate
Baselines: PhotoMI AdaConv WCT QuantArt InST OcTree MedianCut ColorCNN
Builds on: UNeXt ColorCNN reCAPTCHA v2 PhotoMI
Closest papers by meaning
- MCA-Bench: A Multimodal Benchmark for Evaluating CAPTCHA Robustness Against VLM-based Attacks (AAAI 2026 · similarity 0.86)
- Spatial CAPTCHA: Generatively Benchmarking Spatial Reasoning for Human-Machine Differentiation (ICLR 2026 · similarity 0.83)
- CAPTCHA Solving for Native GUI Agents: Automated Reasoning-Action Data Generation and Self-Corrective Training (ICML 2026 · similarity 0.82)
- Variance-Based Membership Inference Attacks Against Large-Scale Image Captioning Models (CVPR 2025 · similarity 0.82)
- Open CaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents (NeurIPS 2025 · similarity 0.82)
- Next-Gen CAPTCHAs: Leveraging the Cognitive Gap for Scalable and Diverse GUI-Agent Defense (ICML 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.