Cowpox: Imaging Diagnosis & AI Defense

• Cowpox is a distinct skin lesion class used in machine-learning studies to differentiate mpox, chickenpox, and other similar diseases.
• It appears in datasets like the Kaggle-based split and MSLD v2.0, emphasizing real-world challenges such as class imbalance and image variability.
• In AI security, 'Cowpox' names a distributed defense strategy that deploys benign cure samples to immunize vision-language models against adversarial attacks.

Cowpox, in the literature considered here, appears in two distinct technical senses. In dermatology-oriented machine-learning studies, it is an explicit skin-lesion class used in multiclass differential diagnosis against mpox and other eruptive diseases, rather than a loosely defined “other rash” category (Iqbal et al., 5 Jan 2026, Ali et al., 2023). In AI security, “Cowpox” is also the name of a defense mechanism for infectious jailbreaks in vision-language-model multi-agent systems, chosen by analogy to immunization (Wu et al., 12 Aug 2025). The dominant biomedical treatment in these sources concerns cowpox as a visually confounding comparator in skin-image classification pipelines, where it is consistently framed as difficult to separate from chickenpox and, to a lesser extent, mpox (Iqbal et al., 5 Jan 2026, Ali et al., 2023).

1. Terminological scope and corpus-defined meaning

Within the biomedical papers, cowpox is represented operationally as a supervised class label in image datasets built for mpox screening or lesion classification. One study uses a five-class benchmark comprising MPox, Measles, Cowpox, Chickenpox, and Normal (Iqbal et al., 5 Jan 2026). Another introduces MSLD v2.0, a six-class dataset containing mpox, chickenpox, measles, cowpox, hand-foot-mouth disease (HFMD), and healthy, with cowpox included as one of the five non-mpox differential diagnosis classes (Ali et al., 2023).

These studies do not provide a full clinical review of cowpox. They do not supply lesion-stage descriptions, morphology-specific diagnostic rules, or systematic clinical differentiation criteria. Instead, they position cowpox as a look-alike lesion class that matters because mpox diagnosis from photographs is difficult when several rash diseases show high inter-class similarity and substantial intra-class variability (Iqbal et al., 5 Jan 2026, Ali et al., 2023). This suggests that, in the current computational literature, cowpox is treated primarily as a problem of visual differential diagnosis rather than as a richly characterized clinical endpoint.

A separate and unrelated usage appears in AI robustness research, where Cowpox names a distributed defense for VLM-based multi-agent systems. There, the term is metaphorical rather than biomedical: a benign “cure sample” is propagated through the system to immunize agents against infectious adversarial samples (Wu et al., 12 Aug 2025).

2. Cowpox in lesion-image datasets

Cowpox is explicitly present in two dataset formulations summarized in these papers. In the five-class Kaggle-based setting, the total dataset is split into Train & validation (80%) with 10,643 images and a Test set (20%) with 3,326 images, using an Input size of 224 × 224 (Iqbal et al., 5 Jan 2026). The paper states that cowpox is a dedicated class and that the class distribution reflects a real-world setup with class imbalance, but it does not provide a clean tabulation of the exact number of cowpox images or the exact cowpox train/validation/test split (Iqbal et al., 5 Jan 2026).

In MSLD v2.0, cowpox has a reported class size of 66 images (Ali et al., 2023). The dataset is retrospective and web-scraped from publicly available case reports, news portals, and reliable websites; all images, including cowpox, undergo quality filtering, cropping to the lesion region of interest, resizing to 224 × 224 while maintaining aspect ratio, and expert dermatologist verification of disease labels (Ali et al., 2023). The split is approximately 70:20:10 for train/validation/test with patient independence preserved, but no classwise split table is provided for cowpox (Ali et al., 2023).

| Dataset/study | Classes including cowpox | Cow