It appears to me that this mostly fools whole-image classification algorithms. If the system performs object segmentation first and then applies classification to each object in the scene, this method is unlikely to be effective.

One can paste such a sticker on top of the face or other objects to be disguised and it might reduce recognition accuracy a bit, but applying some “paint-in” algorithms to fill in the blank covered by the sticker would basically remove its effect. That is unless it is used to cover some prominent features, although that is often unpractical in many circumstances.