Research Interests

Digital Watermarking

Digital watermarking consists in modifying multimedia content (should it be music, image, video, 3D meshes, etc) in an imperceptible manner so as to convey additional information in a robust fashion. In other words, the watermarking process is deemed to introduce some distortion but the introduced modifications should remain imperceptible by a human being. Moreover, the hidden message encoded by these modifications is expected to be retrieved whatever signal processing primitive the signal may experience after watermarking e.g. noise addition, filtering, lossy compression, etc. In particular, the embedded watermark is expected to survive D-A/A-D conversion (print-and-scan, project-and-camcord, playback-and-record, etc) aka. the analog hole. This feature makes digital watermarking an appealing technology to complement conventional cryptography-based protection systems whose protection inherently disappears at rendering time.

Digital watermarking is not a graphic overlay (alpha blending). It is strictly unnoticeable by the human sensory system. Digital watermarking is not steganography. Steganography indeed focuses on statistical undetectability i.e. an automaton should not be able to reliably tell if some content contains a hidden message or not. In contrast, digital watermarking cares about sensorial imperceptibility i.e. a human being should not be able to reliably tell if some content is watermarked or not. Digital watermarking is not (only) data hiding. Data hiding is interested in high data rate auxiliary communication channels in a controlled environment i.e. with no or small known perturbations. On the other side, digital watermarks are expected to survive even crude operations and do not necessarily require to encode a large number of bits.

Digital watermarking was originally intended for copyright protection applications. For instance, audio watermarks are embedded today in some bluRay disks so as to activate relevant playback control policies on compliant players. However, due to known weaknesses against oracle attacks, the scope of watermarking has shifted over the years from piracy prevention to piracy deterrence with applications such as broadcast monitoring to accurately redistribute collected royalties or traitor tracing to pinpoint the source of a leak in a distribution framework. The embedded watermark does not prevent the act of piracy but will reveal it afterwards. Digital watermarking could also be exploited to perform metadata binding and provide enhanced services, such as second screen synchronization for instance.

Content-based Fingerprinting

Content-based fingerprinting is also commonly referred to as passive fingerprinting, perceptual hash, robust hash, or soft hash. It refers to a technology whose objective is to generate a DNA-like binary digest which could be used to identify multimedia content. In contrast with conventional cryptographic hash functions whose output hash values are drastically altered as soon as the input is modified by a single bit, content fingerprints are expected to exhibit some robustness to conventional signal processing primitives such as noise addition, filtering, lossy compression, etc. In other words, altering multimedia content should have a marginal impact on the extracted fingerprint as long as the semantic of the content remains untouched i.e. the fingerprint should gracefully degrade with the quality of the content. In comparison with digital watermarking, content-based fingerprinting does not require any modification of the content prior to distribution and can thus be readily applied to legacy content in order to incorporate already published items into the system.

Due to its inherent properties, content-based fingerprinting is particularly fit for content identification purposes. For instance, it has been heavily popularized with Name That Tune applications such as Shazam. It is also routinely exploited for copyright protection applications such as broadcast monitoring or content filtering e.g. to verify that uploaded user-generated content does not infringe any copyrighted item registered in a reference fingerprint database. Alternatively, content fingerprints could be used as an index in a database to bind relevant metadata and thus provide enhanced services. Finally, content fingerprints could also be exploited to help the resynchronization process of the detector in a watermarking system. For instance, if the content fingerprinting system can provide an accurate temporal index in a movie, it can be exploited to realign the analyzed movie in time. Similarly, accurate localization of feature points in a video frame is most useful to compensate for any spatial distortion e.g. inverse homographic transform to accommodate for camcording attacks.

Multimedia Forensics

Multimedia passive forensics is the last line of defense in copyright protection. What can you tell about a pirated digital item if (i) it has not been watermarked prior to distribution, and (ii) it has not been registered into a fingerprint database? Passive forensics analysis focuses on the very essence of the content in order to reveal its history. What kind of acquisition device has been used to generate the content? What signal processing primitives were applied to the content? To do so, passive forensics analyze the statistics of the signal in order to pinpoint unexpected statistical inconsistencies. Such artifacts are then uniquely associated to a particular primitive or device. These forensics clues are currently envisioned to be most useful to assess the authenticity of digital documents, should it be in court or for press agencies, but they could also be exploited to reverse-engineer the 'workflow' of pirates.

Multimedia forensics is a relatively new research area but it has been receiving increasing interest over the last few years. Most well-known results include verifying that a piece of content has been captured with a particular device using the statistical fingerprint left by a collection of imperfect sensors, detecting lighting inconsistencies, isolating copy-paste forgeries in multimedia content, inferring the position of a camcorder based on the observed projective distortion in the analyzed content, etc. As usual, there is an on-going cat-and-mouse game between forensics techniques designers and adversaries who aim at evading detection. Although it seems like it would be feasible to spoof a single forensic technique, it looks like it would be difficult to defeat a large array of forensics tools without significantly degrading the quality of the content itself.

Content Protection Systems

Content protection systems typically refers to content distribution systems which rely on cryptographic primitives, e.g. encryption, to prevent unauthorized users to have access to protected content. It encompasses both Conditional Access Systems (CAS) in the broadcast arena and Digital Rights Management (DRM) in the broadband arena. Such protection techniques grew steadily in the late 90's but rapidly received negative publicity due to their lack of interoperability. Content protection were historically built in silo, meaning that content protected by one system could not be played on another one. This situation goes against the Pay Once, Play Everywhere expectation of the consumers and therefore triggered increasing frustration since they could not use the content, which they have paid good money for, as they wished.

Throwing bridges between closed-wall content protection systems is not a straightforward task. First, it requires the good will of the DRM and CAS vendors. Second, it inherently induces some content dilution which needs to be tightly controlled in order to preserve the revenues of the copyright holder. Means to reduce content dilution typically reduces today to three simple primitives: counting, timing and measuring distance. Combining these three primitives, constructors were able to provide protection services with increasing interoperability over the last decade, with the apparition of domains and digital rights lockers for instance. Still, a couple of 'historical usages' are still not supported. This include for example the first sale doctrine (it is not possible to resell a protected digital item), gifting, lending, giving away, etc. As a matter of fact, the legality of some of these usages is even questionable in some cases e.g. reselling a digital item in the United States.