Files
Wuillaume_90952100_2026.pdf
Open access - Adobe PDF
- 3.13 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Images are everywhere. On social media, newspaper, on television, … everywhere. Lately, the IA are generating more realistic content than ever. The issue is precisely that is becomes more and more difficult for a human to distinguish which content has been generated from the one that is authentic. The goal of this Master Thesis is then to propose a method that will help at proving the authenticity of a given image. The proposed method allows to first secure an image and then to later prove its authenticity by checking the elements that has been introduced in the securing process. The proposed framework relies on image’s metadata, blind watermarking and classification task, resulting in an authenticity score to an image. As part of the process, the method also computes a RASH signature, a hash signature that is unique to an image, up to a benign transformation. The computation of the signature is based on the features points detected in the image such that it improves the robustness of the original technique. According to the experiments carried out, the RASH signature is robust when it should be, and it also detects malicious manipulation. The metadata is also a robust mean to embed data since it has always been successfully retrieved in the performed experiments. Watermark is a robust data method but it shows some sign of weakness when some attacks has been performed of the original images. Finally, the classifier shows good performance when the prediction is done on unseen images from the training dataset. However, if the prediction is done on a completely new dataset, the performance of the classifier decreases significantly. The classification task is then dataset-dependent. With everything together, the proposed framework has given satisfactory results since the performed set of tests have always return desired results.