Domain Name Homograph Detection Using CNN Visual Similarity

The patent replaces the mechanical/algorithmic approach of integer code point comparison with a neural network-based system that processes visual representations of characters. The CNN model takes image inputs of characters and outputs predictions about their visual similarity, substituting traditional computational methods with a learning-based system that captures visual patterns without relying on numerical codes.

The patent introduces an intermediary layer between the input domain name and the detection decision. Instead of directly comparing integer code points, the system converts characters to visual representations (images) and uses a neural network as an intermediary processor to determine visual similarity. This intermediary step enables more accurate homograph detection while maintaining system manageability through modular architecture.

The patent applies preliminary action by pre-training the CNN model offline using a large dataset of character images and their visual similarities. During this training phase, the network learns to recognize visual patterns and relationships between characters. Once trained, the model can make rapid predictions on new inputs without requiring extensive computational resources, thus reducing real-time processing time while maintaining high accuracy.

The patent utilizes parameter changes by adjusting the model architecture and training parameters to optimize the balance between accuracy and speed. The CNN model incorporates specific layers and activation functions that enable efficient feature extraction. Additionally, the system processes domain names character-by-character, breaking down the complex task into smaller units that can be processed more quickly, thereby reducing overall detection time.