Deep neural network is gullible: accurately identifying difficult-to-distinguish images

Recently, Deep Neural Networks (DNNs) have achieved the most advanced performance in various pattern recognition tasks, most notably the problem of visual classification.

Recently, Deep Neural Network System (DNNs) has achieved advanced performance in recognizing images in various modes, especially in visual classification.

In view of the fact that DNA NSA can now classify the objects in the image with a performance close to human level, the question naturally arises: What are the differences between computer and human vision?

Now DDNs can classify images at the level of human vision, so naturally, people still have differences on the visual differences between computers and humans.

A recent study shows that changing an image in a way that humans can't detect (such as a lion) will cause DNN people to completely mark the image as something else (such as mistakenly marking a lion as a library).

A recent study shows that an image (such as a lion) can be changed in a way that humans can't perceive, but DDNs can recognize the change of the image and mark it as a completely different image (just like a lion mislabeled in a library).

Here, we show a related result: it is easy to produce images that human beings can't recognize at all, but the most advanced DNNs thinks that objects can be recognized with 99.99% confidence (for example, definitely marking white noise as a lion).

Some related research results show that it is obviously difficult for human beings to distinguish images, but the most advanced DDNs can distinguish objects with 99.99% accuracy (for example, it can distinguish lions from static white noise).

Specifically, we use convolutional neural networks that are trained to perform well on ImageNet or MNIST datasets, and then find images with evolutionary algorithms or gradient ascent, which DNNs marks as belonging to each dataset class with high confidence.

Specifically, our convolutional neural network performs well on image networks or MNIST data. However, next, we will find that DDNs can use each data set class to highly accurately identify those images that have undergone evolutionary algorithm and gradient rise.

It is possible to produce images that are completely unrecognizable to human eyes. DNNs almost certainly thinks that these images are familiar objects, which we call "fooling images" (more generally, fooling examples).

It is entirely possible for us to produce images that humans can't recognize at all, but these images are indeed familiar and recognizable objects for DDNs, and we call them "deceptive images" (more common and confusing images).

Our results reveal interesting differences between human vision and current DNNs, and question the universality of DNN computer vision.

Our research results reveal the interesting differences between human vision and current DDNs, and raise people's attention to the universality of computer vision.