Tatouage numérique d'image robuste aux distorsions géométriques Détection de données hors distribution pour la classification

The first part proposes a blind zero-bit image watermarking scheme for the unobtrusive protection of photographic works, designed to withstand geometric distortions and the print/scan (P/S) cycle. The core idea is to operate in a tractable invariant domain: the watermark is inserted and detected via the radial average of the amplitude spectrum, which makes the embedding additive on the magnitude of the discrete Fourier transform and simplifies the entire pipeline. This simplification enables game-theoretic signal shaping, in line with Cox's principle (embedding the watermark in perceptually significant components), to maximize robustness under an imperceptibility constraint. Altogether, this frames an efficient pipeline: selection of a useful frequency band, correlation and thresholding at detection, and demonstrated robustness to geometric and P/S attacks.In the second part, we focus on detecting out-of-distribution (OOD) data in deep learning classification. OOD data are samples drawn from a probability distribution different from the one governing the data used to train the classifier. The classifier's decision on such inputs is not meaningful; they must therefore be detected as such to ensure system reliability. In a first approach, we structure the latent space of an autoencoder to perform both classification and OOD detection by estimating the likelihood of the input given the target classes. To further separate OOD data in the autoencoder latent space, we then propose a method for finetuning the model on the tested batch. The method's performance is evaluated and compared with state-of-the-art OOD detection techniques, and within the broader framework of selective classification with OOD detection (SCOD).