Convolutional neural network (CNN)

A convolutional neural network exploits an assumption a general neural network does not make: that a pattern means the same thing wherever in the picture it appears. Its input is data laid out on a grid, usually an image, and it examines small patches at a time using the same reusable filters, looking for local structure: an edge here, a corner there, a texture. Early layers find simple shapes, later layers combine those into complex ones, and the network works out for itself which patterns are worth detecting. Nobody specifies what to look for. The training itself is ordinary supervised learning: labelled images in, learned filters out.

That self-directed feature learning is what made CNNs dominant for anything visual, and security uses them in more places than people expect: document and identity verification, CCTV analytics, deepfake detection, and malware classification where binaries are rendered as images so a vision model can be pointed at them.

It is also why they are difficult to govern. The features a CNN learns have no names and no guaranteed correspondence to anything a human would consider relevant. A network can reach excellent accuracy by keying on an artefact of how the training images were collected rather than on the thing you meant it to recognise, and you will not discover this by reading the model. The practical consequence is adversarial examples. Because the learned features are not the ones humans use, changes far too small for a person to notice can be calculated to flip the network’s answer entirely. A printed pattern can defeat a detector no human would be fooled by. Any deployment where being wrong matters needs testing against deliberately adversarial input, not only a held-out accuracy score.