StegaPos: Preventing Crops and Splices with Imperceptible Positional Embeddings

Abstract

We present a model that allows differentiating between images that are authentic copies of ones published by photographers, and images that have been manipulated by cropping, splicing or downsampling after publication. The model comprises an encoder that resides with the photographer and a matching decoder that is available to observers. The encoder is trained to embed imperceptible positional signatures into image color values at publication time. The decoder is trained to use these steganographic positional (``stegapos'') signatures to determine, for each small image patch, the 2D positional coordinates that were held by the patch within its originally-published image. Crop, splice and downsample edits are then detectable by the inconsistencies they cause in the hidden positional signatures. We find that training the encoder and decoder together results in a model that preserves visual quality while outperforming existing splice detectors on established benchmarks and achieving high accuracy on a new benchmark for crop detection.

Citation

                    @misc{egri2021stegapos,
                          title={StegaPos: Preventing Crops and Splices with Imperceptible Positional Encodings}, 
                          author={Gokhan Egri and Todd Zickler},
                          year={2021},
                          eprint={2104.12290},
                          archivePrefix={arXiv},
                          primaryClass={cs.CV}
                    }