Cyber Social Security requires efective tools for the identification and automated moderation of harmful visual content, such as non-consensual nudity, sextortion, and online pornography. Addressing this issue requires not only accurate AI-based moderation tools but also systems that align with ethical, trustworthy, and human-centered design principles. In this study, we present a comparative analysis of two versions of the YOLO framework (YOLOv5 and YOLO11), evaluated across their respective model sizes (n, s, m, l, x, xl) and tested with both pretrained and randomly initialized weights. The goal is to determine the most efective configuration for the task of nudity detection. To this end, we constructed a dedicated dataset of over 5,000 annotated images across ten sensitive classes, with a focus on semantic balance and annotation quality. The models were tested under various configurations, revealing that YOLO11m with pretrained weights ofers the best trade-of between accuracy and computational eficiency. The results confirm the potential of YOLO-based models for real-time automated moderation applications, while also highlighting the need for further improvements in localization accuracy.
The development of the internet and digital technologies has both enhanced social interaction and the
availability of damaging material. Concerned Cyber Sociologists need mechanisms capable of automated
non-consensual image-sharing detection and online sexual exploitation deterrence, capable of nudity
detection as it relates to image analysis [
In addition, recent research explores the potential of human-AI symbiosis and human body part
detection for advanced human-machine interaction. Willcox & Rosenberg [
Therefore, considering the need to identify new tools for social security and the literature on nudity detection in social contexts, this paper describes a system for detecting nudity that uses exclusively the YOLO architecture which trains to find and mark nude boundaries in static images. While twostage approaches must be less accurate, focus on inference speed, and greater ease-of-use, single-stage models derive better results. In this study, we attempt to determine which variant of YOLOv5 and YOLOv11 enables real-time moderation of explicit content based on accuracy, eficiency, and resource consumption.
The automated detection of pornographic and sexually explicit content is a central challenge within
the broader field of Cyber Social Security [
Early approaches to visual explicit content detection primarily relied on color-based models to
identify skin-toned regions under various lighting and pose conditions [
To improve robustness, mid-level representations such as the Bag of Visual Words (BoVW) were
introduced, combining local feature descriptors with classifiers like SVMs for enhanced
discrimination [
A major leap occurred with the advent of deep learning, particularly Convolutional Neural Networks
(CNNs). AGNet, an ensemble of AlexNet and GoogLeNet, achieved 89.2% accuracy on the NPDI dataset
by aggregating predictions across frames [
More recently, the focus has shifted toward computational eficiency and real-time deployment. Mallmann et al. [18] introduced PPCensor, a CNN-based pipeline that reframes nudity detection as an object detection task. By applying localized obfuscation to private body regions, the system allows for granular moderation without discarding entire frames, while maintaining near real-time performance on edge hardware.
In parallel, transformer-based architectures have gained attention for their ability to capture global context. He et al. [19] demonstrated that Vision Transformers (ViTs) significantly outperform traditional CNNs such as ResNet in classifying sensitive content, thanks to their self-attention mechanisms.
YOLO-based methods have also emerged as promising alternatives for adult content detection. Typically, these systems follow a two-stage architecture: first detecting people or sensitive body parts using YOLO, followed by a secondary classification network [ 20, 21]. While efective, this separation introduces architectural complexity and additional inference latency.
Our work departs from this paradigm by employing YOLO in a fully end-to-end manner. We train
the network directly to detect explicit regions without auxiliary classifiers, resulting in a single-stage
architecture that reduces latency and simplifies deployment—particularly in real-time applications.
Unlike prior video-based methods that apply naive frame-by-frame processing [
YOLO (You Only Look Once) is a unified, real-time approach to object detection proposed by Redmon et al. (2016) [22], which reformulates the detection problem as a single regression task that directly maps from image pixels to bounding box coordinates and class probabilities.
The architecture of YOLO is based on a unified convolutional neural network that processes the entire image in a single pass. The image is divided into a grid of size × , where each cell is responsible for detecting objects whose center falls within it. Each cell predicts bounding boxes, each with a confidence score that reflects both the probability of the presence of an object and the spatial accuracy of the prediction, calculated using the Intersection Over Union (IoU) metric. In parallel, each cell provides a single conditional probability distribution over the classes, which is computed only if the cell contains an object.
YOLO was chosen for its: • Speed – since it treats the problem as a regression task, it does not involve a complex pipeline; • Contextualization ability – as it has a global view of the image during both training and testing; • Generalization capability – as it learns generalized representations of objects. 3.1. YOLOv5 YOLOv5 incorporates the Cross Stage Partial Network (CSPNet) [23] into its backbone (Darknet). CSPNet reduces redundant gradient information during training, thereby improving the model’s eficiency. It splits the feature map into two flows: one is processed through a series of convolutional blocks, while the other remains unchanged. In the end, the two flows are concatenated, reducing the overall number of parameters and computational cost (in terms of FLOPs), without compromising performance.
In the neck, the model adopts the Path Aggregation Network (PANet) [24], which enhances information transmission between diferent levels of the network by adding a bottom-up path to the traditional top-down structure of the Feature Pyramid Network (FPN). This enables better propagation of both low- and high-resolution features, contributing to more accurate object localization.
Finally, the head of the network consists of three convolutional layers. The activation functions used are SiLU and Sigmoid: the former is applied in the hidden layers, while the latter is used in the output layer. The model outputs three types of predictions: the classes of the detected objects, their bounding boxes, and their objectness scores. The CIoU (Complete Intersection over Union) is used to compute the location loss. 3.2. YOLO11 YOLO11 [25] represents a significant advancement of the YOLO framework. The main innovations introduced in YOLO11 include: • C3k2 Block: a more eficient variant of the classic CSP Bottleneck module. It uses two convolutions with smaller kernels instead of a single larger one, reducing computational cost while maintaining good performance. Its behavior can vary based on the c3k parameter, allowing for deeper structures when needed. • C2PSA Block: introduces a spatial attention mechanism that helps the model focus more efectively on the most relevant areas of the image, improving detection accuracy, especially in complex scenes or with small or partially occluded objects. • CBS Blocks (Convolution-BatchNorm-SiLU): combine convolution, batch normalization, and SiLU activation to enhance the quality of the extracted features, making the learning process more stable and efective, and contributing to greater accuracy.
With respect to traditional YOLO architecture, the innovations introduced are arranged as follows: • Backbone: replacement of the C2f block with the more eficient C3k2, retention of the SPPF block, and introduction of the new C2PSA to enhance spatial attention. • Neck: use of the C3k2 block to improve speed and reduce computational complexity, along with integration of the C2PSA block to increase the relevance of features, especially for dificult-todetect objects. • Head: combined use of C3k2 and CBS blocks to process feature maps and increase detection accuracy. This section ends with 2D convolutional layers and the Detect module, which produces the final output (bounding boxes, confidence scores, and classes). The behavior of the C3k2 block is governed by the c3k parameter, which adjusts its internal structure.
For both YOLO versions, the Ultralytics platform ofers model implementations in two configurations: • one pre-trained on the COCO dataset; • another initialized with randomly assigned weights.
The primary objective of our experimental evaluation is to identify the most suitable YOLO model variant for end-to-end nudity detection in static images. To this end, we systematically investigate how architectural variations within the YOLO family afect both detection accuracy and computational eficiency.
For the development of our automatic human body part detection system, a dedicated dataset was constructed to address the specific requirements of the task. The dataset was developed through an iterative pipeline comprising repeated cycles of web-based image collection, manual annotation, and empirical evaluation of model performance. Particular attention was given to enhancing dataset quality and coverage through successive refinement steps, which included targeted augmentation of underrepresented classes and exclusion of low-quality or ambiguous samples. This process allowed for the progressive improvement of the dataset in terms of both class balance and semantic diversity.
The final version of the dataset, employed for training the YOLOv5 and YOLOv11 object-detection models, consists of 5 090 images annotated with 8 247 bounding boxes. While the total number of samples remains relatively limited, the dataset reflects a considerable investment of time and manual efort, and its composition was carefully curated to optimize the training process for the intended detection task.
The dataset includes annotations for the following ten classes, encompassing both anatomical features and sexually explicit content: anus, breast, buttocks, penis, vagina, oral-sex, penetration, penetration position, masturbation, porn.
Annotations were performed using bounding boxes in accordance with a consistent labeling protocol designed to ensure inter-annotator agreement and reduce noise in the training data. Class frequencies were regularly monitored throughout the dataset-construction process, and specific measures were taken to mitigate class imbalance and prevent model bias. The resulting dataset thus provides a task-specific and well-structured foundation for the supervised training of explicit-content detection systems.
The evaluation focuses on a comparative analysis of multiple YOLO architectures, emphasizing both the well-established YOLOv5 family and the more recent YOLO11 series. The aim is to determine the optimal model configuration that balances detection performance with computational eficiency for the specific task of nudity detection.
The following model configurations were evaluated: • YOLOv5: YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x • YOLO11: YOLO11n, YOLO11s, YOLO11m, YOLO11l, YOLO11x
• Pre-trained weights from the COCO dataset; • Random weight initialization.
Training was conducted for 100 epochs using the default hyperparameters provided by the respective implementations. All experiments employed consistent data augmentation strategies and loss functions. Input resolution and batch size were adapted per model to optimize GPU utilization while maintaining experimental comparability.
This setup allows for: • Comparative analysis across lightweight, mid-sized, and high-capacity models. • Identification of the YOLO variant ofering the best trade-of between detection performance and computational eficiency.
To ensure systematic monitoring and reproducibility, we utilized Weights & Biases and Comet throughout the training and evaluation phases. These tools enabled comprehensive tracking of: • Precision, recall, and mAP over training steps; • Learning curves and loss values; • All relevant training hyperparameters.
This experimental protocol supports a fair, reproducible, and well-documented comparison of YOLO models of varying complexity under realistic deployment conditions.
This section presents the comparative evaluation of YOLOv11 (Table 1) and YOLOv5 (Table 2) architectures on the task of nudity detection, performed on a curated dataset of over 5,000 annotated images across ten sensitive semantic classes. The goal was to assess the detection capabilities of each model variant, considering both pretrained and randomly initialized configurations, and to identify the optimal trade-of between accuracy and computational eficiency in view of real-time human-AI collaborative applications.
Among the YOLOv5 variants, YOLOv5x achieved the best results, with a mAP@0.5 of 0.367 and mAP@0.5:0.95 of 0.215. Precision and recall were 0.412 and 0.440 respectively, reflecting a reasonably balanced detection performance. Smaller configurations such as YOLOv5n and YOLOv5s showed a significant drop in recall, limiting their applicability in critical moderation tasks. The inclusion of pretrained weights yielded modest improvements across all sizes. mAp50 mAP50- Precisio 95 n
Recall
In contrast, YOLO11 models consistently outperformed YOLOv5 in both accuracy and recall. The YOLO11m configuration achieved the best results overall, with a mAP@0.5 of 0.438, mAP@0.5:0.95 of 0.243, and a recall of 0.516, outperforming all other configurations. Notably, models trained from scratch showed a marked decrease in performance—e.g., YOLO11m without pretrained weights reached only 0.291 in mAP@0.5—highlighting the importance of transfer learning, particularly in domain-specific visual tasks such as nudity detection.
A direct comparison between YOLOv5x and YOLO11m, summarized in Table 3, demonstrates the superior capability of YOLO11m, especially in detecting nuanced and sensitive content. These results underscore the potential of YOLO11-based architectures to support automated moderation systems that are both accurate and eficient.
From a human-AI symbiosis perspective, high recall and precision rates are essential to ensure user trust, system transparency, and ethical alignment. YOLO11m’s performance enhances the reliability of AI-based moderators in identifying harmful visual content, minimizing both false positives and false negatives. Furthermore, the adaptability shown by pretrained configurations supports future personalization and domain transfer, critical for sensitive contexts such as healthcare, education, or platform moderation.
Extensive empirical evaluation revealed that the Medium (M) and Large (L) configurations of the YOLO architecture demonstrated the most favorable performance for human-body-part detection, particularly when initialized with pre-trained weights. These configurations ofered an optimal compromise between detection accuracy and computational eficiency, rendering them appropriate for deployment in practical applications such as online-safety monitoring and content moderation.
Nonetheless, as illustrated by the visual results, the models’ overall detection performance remained suboptimal. Although the networks exhibited a capacity to identify relevant anatomical features, they mAp50 mAP50-95 Precision
Recall Train/ Train/ Train/ Val/box box_lo cls_lo dfl_lo _loss ss ss ss
Val/cls_l Val/df oss l_loss
YOLOv5n YOLOv5n _nw YOLOv5s YOLOv5s _nw YOLOv5 m YOLOv5 m_nw YOLOv5l YOLOv5x _nw YOLOv5l_ 0.34984 nw YOLOv5x 0.36713 frequently encountered dificulties in achieving precise object localization and accurate delineation of bounding boxes. Among all evaluated variants, the YOLOv11-M model with pre-configured weights proved to be the most efective, yielding the highest precision scores. However, it still exhibited notable shortcomings in terms of boundary accuracy and spatial consistency.
These observations suggest that, while YOLO-based models hold promise for the task of body-part detection, further enhancements—such as more meticulous annotation, the incorporation of additional training samples, or architectural refinements—are required to improve localization precision and overall detection robustness.
This work was partially supported by the following projects: SERICS - ”Security and Rights In the CyberSpace - SERICS” (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU; Patto territoriale "Sistema universitario pugliese" – CUP F61B23000370006; Accordo Quadro CrASte - “Cyber Academy for Security and Intelligence”.