Artificial intelligence-based face recognition solutions are becoming increasingly popular. Therefore, it is crucial to fully understand and explain how these technologies work in order to make them more effective and acceptable to society. This is the goal of the CHIST-ERA project XAIface, the final results of which are reported in this article: a framework and toolkit for improving AI decision explainability, in the context of automated face recognition, through several novel methods are presented. These methods are integrated into an end-to-end face recognition demonstrator system, which facilitates studying the impact of various influencing factors and system processes on recognition performance. By doing so, we can visually explain the decisions made by the face verification pipeline for specific instances in our test set using heatmaps and locally interpretable features. Furthermore, we offer a comprehensive explanation of the end-to-end model by examining the relationship between verification failures and misclassifications of soft biometric facial traits.

Recognition of the social network of origin of an image is a relatively recent topic that is part of the techniques that fall under the umbrella of digital image forensics. It consists of the classification of images according to the social network on which they were posted. In contrast with other topics of digital image forensics, there are no works addressing counter forensic for source social network identification. Thus, we analyse the impact of image manipulations on its performances. We focus our study on AI-based compression, which tends to become the new compression solution with the upcoming standard JPEG AI. To conduct a fair analysis, we compare the AI-based compression with the conventional legacy JPEG compression, and also include three other manipulations: median filtering, Gaussian blurring, and additional white Gaussian noise, which are often used to assess the robustness of digital image forensic methods. We define two sets of parameters based on the resulting image quality in terms of structural similarity, which correspond respectively to attacks with strong and limited image degradation. In the context of strong downgrade of the image quality, all the manipulations lead to similar decrease in performance, while for attacks that preserve image quality, AI-based compression is able to reach a drop in identification rate twice higher than the other manipulations.

International audience

The identification of the source social network from an image is a relatively new research area in the image forensic domain. The classification of the source social network can be a crucial element for the growing number of cases of social-media related crimes, such as cyberbullying. This paper takes into consideration the state-of-the-art approaches addressing this problem and proposes a new methodology to improve the results obtained to date. Our identification technique is based on the idea that social networks perform some processing on the uploaded images, such as resizing or recompression, and leave some artifacts on them. We propose to use discrete cosine transform features and image noise residual analysis to detect such artifacts. A two-stream convolutional neural network, which combines the inputs from these two artifact domains, is trained to classify the source social network of images coming from three different datasets. This paper explores the two domains, proposes strategies for managing unbalanced datasets, provides details about the proposed two-stream convolutional neural network, and presents the results achieved by our method compared with the current state-of-the-art approaches.

The paper explores how light fields captured by plenoptic cameras can increase the performance of face landmark detection. The idea is to exploit light fields geometrical constraints to correct the position of points detected by classical face landmark detectors. These geometric constraints are used to enforce landmark points angular coherency across the different views of the light field, and by doing so to correct the positions of the landmarks on all views. The corrected landmark points are compared with ground-truth manual annotations of a set of 400 images corresponding to the central views of 400 light fields of faces with different pose and expression.

A novel approach for fast iris recognition on mobile devices is presented in this paper. Its key features are: (i) the use of a combination of classifiers exploiting the iris colour and texture information; (ii) its limited computational time, particularly suitable for fast identity checking on mobile devices; (iii) the high parallelism of the code, making this approach also appropriate for identity verification on large database. The proposed method has been submitted to the Mobile Iris CHallenge Evaluation II. The test set employed for the contest evaluation is made available on the contest web page. The latter has been used to assess the performance of the proposed method in terms of Recognition Rate (RR) and Area Under Receiver Operating Characteristic Curve (AUC).

In this paper an iris detection scheme for noisy images acquired by means of mobile devices is presented. Iris segmentation is accomplished by exploiting the use of the watershed transform with the purpose of identifying the iris boundary as much precisely as possible. After a pre-processing step aimed at color/illumination correction, the watershed transform is computed and suitably binarized. Circle fitting is then accomplished to identify the limbus boundary by using curvature approximation and a cost function for circle scoring. The watershed transform is furthermore employed to distinguish, in the zone delimited by the best fitting circle, the regions actually belonging to the iris from those belonging to eyelids and sclera. Finally, pupil detection is accomplished by means of circle fitting and by using a voting function based on homogeneity and separability criteria. The suggested iris detection scheme has a positive impact on an the accuracy in computing the iris code, which has in turn a positive impact on the performance of iris recognition.