In this paper, an unsupervised learning-based autoencoder and decoder are used to find image forgery detection. People access internet and post images on social media sites. This work provides social security by recognizing forged images on online social media sites. Images are processed with colour illumination and converted into positive and negative patches. These patches are stored in a.npy array of (30 × 30 × 3) sizes. The Autoencoder train true positive and true negative patches. An auto decoder reconstructs images from the minimum most essential pixels. The auto-encoder uses a nonlinear transformation to reduce the number of dimensions. The color-illuminated images were applied to Harrie's corner detector machine learning. First, it encodes the input into simple signals. Next, it comprises multiple convolution layers followed by an adder from output three and output 4 with max Pooling. Then, it down-samples the input image up to the maximum point of compression. An auto decoder is used to reconstruct images from the minimum most essential pixels. It can replicate the output image into an input image with some degraded quality. It comprises multiple convolution layers followed by an adder from output three and output four with upsampling. Experiments were performed on various openly available databases like CASIA v1.0, BSDS300, COMOFOD and CASIA v2.0.