Novosibirsk State Technical University Scientists Train Neural Network to Distinguish Real Images From AI-Generated Ones
Specialists at Novosibirsk State Technical University have developed a method that automatically determines whether an image is a real photograph or was created by a neural network.
Distinguishing synthetic images from real ones is becoming increasingly difficult, posing risks of spreading fake content, falsifying evidence and violating copyright. The researchers compared different algorithms and found that the best results came from an approach based on Vision Transformer, which analyzes an entire image rather than individual fragments.
Testing showed an accuracy of 89%. The method does not require tens of thousands of labeled examples and does not lose accuracy when images are compressed. It is planned for use in content verification systems for media outlets, social networks and law enforcement agencies.
