How Wukong’s AI Porn Detection System Achieves 99.5% Accuracy

Image Classification and Porn Detection

Image classification is a fundamental computer‑vision task; the Wukong porn‑detection system treats it as a three‑class problem: porn (explicit sexual content), sexy (non‑explicit but suggestive), and other .

Challenges Specific to Porn Detection

Unlike standard datasets such as ImageNet, porn detection faces:

Multi‑label data where images may contain both porn and sexy elements, requiring a priority order (porn > sexy > other).

Non‑iconic images from real‑world scenes, where tiny regions may contain illicit content.

Severe class imbalance: porn and sexy images occupy a tiny pixel space compared to the vast majority of benign images.

Evolution of Convolutional Neural Networks

Modern image‑classification relies on CNNs. Key milestones include:

AlexNet (2012) – first deep CNN to win ILSVRC.

Network‑in‑Network (2014) – introduced 1×1 convolutions and global pooling.

GoogLeNet (2014) – popularized parallel convolution paths.

ResNet (2015) – residual blocks that balance depth and trainability.

DenseNet (2016) – dense connectivity for feature reuse.

Xception (2017) – depthwise separable convolutions.

For porn detection, a ResNet‑based backbone with custom optimizations proved the best trade‑off between performance and training difficulty.

Transfer Learning

Starting from ImageNet‑pretrained models, only the high‑level layers are fine‑tuned on limited porn‑specific data, leveraging shared low‑level features (edges, textures) while adapting semantic representations to the target task.

Model Visualization Techniques

To understand and improve the model, several visualization methods are used:

Deconvolution (Zeiler, 2013) – maps activations back to image regions.

Activation Maximization – synthesizes inputs that strongly activate a neuron.

Class Activation Maps (CAM) – highlights image areas contributing to a specific class prediction.

Loss Functions

Various loss functions are explored, including Euclidean loss, contrastive loss, hinge loss, sigmoid cross‑entropy, softmax with loss, and TripletLoss, each suited to different aspects of classification and metric learning.

Adversarial Training and GANs

Adversarial examples expose model vulnerabilities; incorporating adversarial training improves robustness. Generative Adversarial Networks (GANs) are employed for:

Generating synthetic porn/sexy samples to enrich scarce data.

Semi‑supervised learning, where GAN‑generated pseudo‑samples help refine decision boundaries.

Image‑to‑Image Translation for Data Augmentation

Pix2Pix and Cycle‑GAN are used to create paired or unpaired image translations (e.g., removing mosaics or converting clothed to nude appearances) to strengthen the model’s ability to distinguish subtle differences between porn and sexy content.

Practical Impact

Deployed on JD.com, the system processes tens of millions of daily requests for product images, user‑generated photos, and public‑cloud content, achieving >99.5% accuracy and reducing manual review workload by over 90%.

Future Directions

Continued exploration of advanced network architectures, better visualization tools, and more sophisticated generative models aims to further improve accuracy, efficiency, and reliability.