Annonce

PhD open position:
Title: Video content security in a deep learning coding architecture
Keywords: video coding, deep learning, security of video content, confidentiality, integrity

Description:
Over the past few decades, numerous video compression algorithms have been developed, most based on a hybrid architecture combining transform coding and predictive coding. Standards such as H.264/AVC, HEVC, and VVC follow this principle. While they offer highly efficient compression performance, each module relies on a rigid, manual design. Furthermore, these modules cannot be jointly optimized end-to-end.

In parallel, recent years have seen the resounding success of deep learning in many disciplines, particularly computer vision and image processing. Consequently, coding architectures based on deep learning and end-to-end optimization have been proposed [Ding 2021, Li 2021, Quach 2022, Chen 2025]. Notably, several solutions have demonstrated competitive performance for video coding compared to traditional approaches.

In this emerging context of deep learning-based video coding, the objective of this thesis is to study the security of video content, particularly its confidentiality and integrity. Although solutions exist within the context of classical encoders [Dufaux 2008, Shahid 2011, Shahid 2014, Boyadjis 2017], to our knowledge, their application to these new encoders has not yet been explored in the literature and raises new challenges.

Initially, to preserve video confidentiality, we plan to study the encryption or obfuscation of variables in the latent space, after quantization but before the entropic encoder. For this purpose, we can consider intra data, residual data, motion data, or a combination thereof. This approach guarantees that the compressed binary stream can still be decoded, but with a noisy reconstructed video. To avoid a significant increase in throughput, it is essential to preserve the statistics of the latent variables. Since the latent space contains semantic information about the content, this approach has the potential to enable selective encryption of certain objects in the scene, such as blurring faces in a video surveillance scenario. In a second phase, we plan to explore content integrity verification. More specifically, we envision using a hash function in the latent space, combined with a digital signature. An attack is detected when the digital signature is missing or when the hash value differs from the value decrypted from the compressed stream.

Starting date: Oct. 1, 2026

Collaboration between:
L2S, CentraleSupelec, Univ. Paris-Saclay (Frederic Dufaux & Giuseppe Valenzise)
CRIStAL, Lille (Vincent Itier)
LIRMM, Montpellier (William PUECH)

For more details and to apply: https://adum.fr/as/ed/voirproposition.pl?matricule_prop=70411&site=PSaclay