Checking Image Integrity for Anticipating Influence Attacks – CI2(IA)

• Detection and localization of deepfakes (IMTNE/CRIStAL)
• Realistic dataset for detector evaluation (GIPSA-LAB)
• Improving detector generalization (IMTNE/CRIStAL)
• Ethical watermarking and generative steganography (IMTNE/CRIStAL)
• Analysis of semantic modifications (IMATAG, IMTNE)
• Analysis of information leaks (IMTNE)

DJIN: Detector based on noise component preservation, avoiding initial pooling layers. Pre-trained on ImageNet for steganography, DJIN is the best detector for in-distribution datasets and is effective on high-quality images of varying sizes.
DinoLizer: Model based on DINOv2 to locate manipulated regions in generative images. It uses a linear classification head on Vision Transformer patch embeddings and a sliding window strategy for large images. DinoLizer outperforms the state of the art with a 12% higher IoU and remains robust to post-processing.
Latent Space Insertion (Glow): The idea here is to use a reversible model, i.e., models that can recover the latent space used to generate the image.
Latent space insertion (Glow): The idea here is to use a reversible model, i.e., models that can recover the latent space used to generate an image
from that image. Although these models are reversible, they do not take into account the fact that an image is not recorded (stored and shared) with floating point numbers. We have
therefore proposed an optimization method to insert a hidden message into the latent space of an image generator. A bias vector learned by gradient descent is added to the latent vector to guarantee a zero error rate after image quantization.
Error reduction through latent space adjustment: This method modifies a high-capacity binary spread spectrum scheme. It corrects errors via an offset assumption and adjusts the latent vector norm to minimize detectability.
RealAI, a highly photorealistic AI-generated dataset designed to provide a challenging evaluation framework for AI-generated image detectors. We evaluate existing detection methods on RealAI. Unlike previous evaluations, we consider not only generalization to unknown generators, but also to images belonging to unknown semantic categories.

• Generative steganography:
Improving the reversibility of diffusion models. The goal is to find a precedent
for the diffusion steps instead of using approximate inversion (i.e., DDIM).
A posteriori error correction. The goal is to use the uniqueness of the image generated for a given message and prompt to ensure that the decoded message is indeed the message transmitted.
• Deepfake localization:
We are currently developing a generative model inpainting dataset that takes into account different models, including new models such as SD3, Flux, and Qwen, in order to refine the Dinolizer model. The goal is to verify performance in real-world scenarios, particularly in an out-of-distribution detection setting.
• Improving detector generalization:
One avenue of research we are currently exploring involves adapting out-of-distribution data, or targets, to training data, or sources, using a process of “whitening and recoloring.” By whitening, we take the training sample and project it into a latent space where all features are decorrelated from each other. In the current simulations, we seek to determine whether VAEs can whiten and recolor data when the alignment is a convolution with a simple filter, which is numerically and mathematically manageable.

M. T. Doi, J. Butora, V. Itier, J. Boulanger, and P. Bas. “DeepFake Detection based on Noise Residuals”, GRETSI, (2025).
M. T. Doi, J. Butora, V. Itier, J. Boulanger, and P. Bas. “DinoLizer: Learning from the Best for Generative Inpainting Localization”, en soumission (2025).
G. Evennou, V. Chappelier, and E. Kijak. “Fast, Secure, and High-Capacity Image Watermarking with Autoencoded Text Vectors”. arXiv preprint arXiv:2510.00799 (2025).
G. Evennou, V. Chappelier, E. Kijak, and T. Furon. “Swift: Semantic watermarking for image forgery thwarting”. 2024 IEEE International Workshop on Information Forensics and Security (WIFS). IEEE. 2024, pp. 1–6.
G. Evennou, A. Chaffin, V. Chappelier, and E. Kijak. “Reframing Image Difference Captioning with BLIP2IDC and Synthetic Augmentation”. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE. 2025, pp. 1392–1402.
I. Messadi, G. Cervia, and V. Itier. “Image selective encryption analysis using mutual information in CNN based embedding space”. 13th European Workshop on Visual Information
Processing (EUVIP). 2025.
A. Noirault, T. Pevny, J. Butora, V. Itier, and P. Bas. “ERROR REDUCTION FOR GEN-AI STEGANOGRAPHY BY ADJUSTMENTS IN THE LATENT SPACE”, en soumission (2025).
A. Noirault, J. Butora, V. Itier, and P. Bas. “Génération d’Images et Insertion d’un Message Caché dans l’Espace Latent de Glow”. Gresti. (2025).
J. L. Thai Ngoc Toan Truong and K. Wang. “Toward generalizable AI-generated Image detection with a new realistic dataset: Performance evaluation and improvement”. IEEE Multimedia. (2026).

The availability of deepfake generation tools makes their creation and dissemination accessible. Moreover, the performance of these models produces results that can deceive humans. In this context, the project Checking Image Integrity for Anticipating Influence Attacks (CI2(IA)) aims at verifying the trustworthiness of visual information coming from journalists, social networks or foreign influences. This project tackles the verification of content integrity, i.e. the ability to determine if a media is a deepfake or not and if it has a malicious purpose.
Although efficient detectors based on neural networks exist, they are only marginally generalizable to data from different distributions (e.g., a different deepfakes generator). To address this issue, domain adaptation techniques will be used in this project.
Another active approach to verify the integrity of deepfakes is to generate “ethical” deepfakes in which a mark is inserted as a watermark to indicate their provenance. The goal of the project is to propose a robust watermarking method, with guarantees on the quality of the generated image, while allowing any user to verify the provenance of a watermarked image.
Active approaches are also used to verify the integrity of watermarked images. By comparing the watermarked image to its modified versions, the project is focused on understanding, automatically, the purposes and implications of semantic modifications between the images.
CI2(IA) proposes to address these complementary aspects in order to provide a critical reading of the information spread in a global context of fake news.