Abstract: An artificial neural network based semi-fragile watermarking scheme for image authentication is proposed.Firstly，some pixels are selected randomly from the host image.Then，the neural network technique is utilized to model relationships between the randomly selected pixels and their 3×3 neighboring pixels.By applying the exclusive-or（XOR） operation to the modeled relationships and the binary watermark image，a secret key for watermark recovery is obtained as the constructed zero-watermark.Since only the image features are extracted from the host image to construct watermark instead of embedding information to the image，avoiding image distortion resulting from watermark embedding.The proposed scheme achieves tamper detection as well as tampered position localization，in addition to the image credibility authentication，while being robust to JPEG compression at a certain degree.The experimental results demonstrate the effectiveness of the proposed scheme.