Abstract: For authenticating an image,this paper proposes a neural network-based fragile watermarking scheme.In the proposed method,we select some pixels and their vicinities randomly,and then establish the relational model among these pixels and other pixels in their vicinities using the neural network.Finally we embed a bit information of the watermark by adjusting the polarity between a selected pixel and the output value of the model.According to the extracted watermark,we can authenticate the tested image and locate the illegal modification. Experimental results show the proposed method has some advantages such as security,high sensitivity and exactly locating the modifications,while still keeping high quality of the image.