Réunion

Impact of Deep Learning on Facial Data Security

Pr. Andréas UHL, Salzburg University, Austria

In face recognition, several techniques have been developed over the years to protect facial templates as well as facial sample data. With the rise of deep learning techniques, face recognition as seen a revolution in terms of achievable recognition accuracy. These developments also had impact on the security techniques applied to facial data. We will demonstrate that early template protection schemes do not provide security any longer in case deep learning based recognition is used. Further, the security of existing selective encryption schemes developed for facial data has to be re-considered in case deep learning based recognition is used. Finally, we will show that learning-based inpainting schemes can be used to successfully attack selectively encrypted facial data.

Partage d'images JPEG secrètes

Pauline PUTEAUX, CRIStAL, Lille

Résumé : De nombreuses méthodes de sécurisation des images JPEG ont été proposées pour lutter contre les menaces sur Internet. En particulier, des méthodes de chiffrement ont été spécialement conçues pour sécuriser visuellement les images JPEG : on parle alors de "crypto-compression". L'inconvénient des méthodes de crypto-compression est de ne dépendre que d'une seule clé secrète. Dès lors, si cette clé est perdue, la totalité du contenu de l'image originale secrète l'est également. Dans nos travaux, nous avons proposé une approche de partage d'images JPEG secrètes. Lors de la compression JPEG, le schéma de partage de secret de Shamir sur les corps de Galois est utilisé pendant l'étape de codage de Huffman. Cela garantit la sécurité visuelle de l'image secrète dans le domaine compressé, tout en résolvant le problème de la perte de la clé secrète. Notre approche est entièrement conforme au format JPEG et préserve la taille par rapport à une compression JPEG standard d'une image originale secrète.

Draco 3D object crypto-compression

Bianca JANSEN van RENSBURG, LIRMM, Montpellier

Résumé : 3D objects have come to play an essential role in industry, however they can also be very large and therefore it is vital that they are compressed. The Draco 3D object compression system, proposed by Google, is becoming an industry standard for 3D object compression. These 3D objects are also important assets which also need to be secured. We propose the first 3D object crypto-compression method, which integrates an encryption step into Google's Draco compression scheme. Our proposed Draco 3D object crypto-compression scheme is format compliant, has no size expansion and does not require additional information such as an auxiliary file.

Can printable unclonable codes be copied? Evaluating the performance of attacks and highlighting the role of the detector

Iuliia TKACHENKO, LIRIS, Lyon

Résumé: Globalization and improvements in digital printing and digitalization technologies have made counterfeiting more prolific and easier to perform than ever. Counterfeits affects the world economics and human life as they are presented in majority of industries (from pharmaceutical, food and agricultural products to valuable documents). Counterfeiting and forgery continue to proliferate partly due to the limitations of existing anti-counterfeiting technologies. One of the promising solutions is the use of printable unclonable codes ? small maximum entropy images, that take full advantage of information loss principle during printing and capturing process. These anti-counterfeiting solutions are successfully commercialized by several leading French and Swiss companies. However, recently the vulnerability of anti-counterfeiting solution based on these printable unclonable codes by deep learning attacks was shown. In short, the apparently simple question: are printable unclonable codes secure against copy?, remains unanswered as of today. In this talk we will highlight the role played by the printable unclonable code detector and its different processing steps. Indeed, depending on the specific processing involved, the detection performance can widely outperform the printable unclonable code bit error rate which has been used as a reference metrics in the prior art.

Engagements flous non linéaires basés sur les codes de Kerdock

Patrick LACHARME, GREYC, Caen

Résumé : les schémas d'engagement flou (fuzzy commitment) sont utilisés pour protéger les données biométriques. Ils sont basés sur des codes correcteurs linéaires, ce qui les rend vulnérables à des attaques par indistingabilité. Dans cet exposé on étudiera comment utiliser des codes non linéaires pour se protéger contre ces attaques, en se focalisant en particulier sur le cas des codes de Kerdock.