Invisible Image Watermarks Are Provably Removable Using Generative AI

Generative AI is the current hot topic. Of course, one of the newest challenges is to discriminate a genuine image from a generative-AI-produced one. Many papers propose systematically watermarking the generative AI outputs.

This approach makes several assumptions. The first one is that the generator is actually adding an invisible watermark. The second assumption is that the watermark survives most transformations.

In the content protection field, we know about the validity of the second assumption. Zhao et al., from the University of California Santa Barbara and Carnegie Mellon University, published a paper. The system adds Gaussian noise to the watermarked image and reconstructs the same image using the noise image. After several iterations, the watermark disappears. They conclude that any watermark can be defeated.

This is a well known fact in the watermark community. The Break Our Watermark System (BOWS) in 2006 and the BOWS2 in 2010 demonstrated this reality. These contests aimed to demonstrate that attackers can defeat the watermark if they have access to an oracle watermark detector.

Thus, this paper illustrates this fact. Their contribution adds generative AI to the attacker’s toolset. As a countermeasure, they propose to use a semantic watermark. The semantic watermark changes the image but keeps its semantic information (or at least some). This approach is clearly not usable for content protection.


Zhao, Xuandong, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. “Invisible Image Watermarks Are Provably Removable Using Generative AI.” arXiv, August 6, 2023.

Craver, Scott, Idris Atakli, and Jun Yu. “How We Broke the BOWS Watermark.” In Proceedings of the SPIE, 6505:46. San Jose, CA, USA: SPIE, 2007.

“BOWS2 Break Our Watermarking System 2nd Ed.”

Watermarking Deep Neural Networks

Recently, an IBM team presented at ASIA CCS’18 a framework implementing watermark in a Deep Neural Network (DNN) network. Similarly, to what we do in the multimedia space, if a competitor uses or modifies a watermarked model, it should be possible to extract the watermark from the model to prove the ownership.

In a nutshell, the DNN model is trained with the normal set of data to produce the results that everybody would expect and an additional set of data (the watermarks) that produces an “unexpected” result that is known solely to the owner. To prove the ownership, the owner injects in the allegedly “stolen” model the watermarks and verifies whether the observed result is what it expected.

The authors explored thee techniques in the field of image recognition:

  • Meaningful content: the watermarks are modified images, for instance by adding a consistently visible mark. The training enforces that the presentation of such visible mark results in a given “unrelated” category.
  • Unrelated content: the watermarks are images that are totally unrelated to the task of the model; normally they should be rejected, but the training will enforce a known output for the detection
  • Noisy content: the watermarks are images that embed a consistent shaped noise and produce a given known answer.

The approach is interesting. Some remarks inherited from the multimedia space:

  • The method of creating the watermarks must remain secret. If the attacker guesses the method, for instance that the system uses a given logo, then the attacker may perhaps wash the watermark. The attacker may untrain the model, by supertraining the watermarked model with generated watermarks that will output an answer different from the one expected by the original owner. As the attacker has uncontrolled, unlimited access to the detector, the attacker can fine tune the model until the detection rate is too low.
  • The framework is most probably too expensive to be used for making traitor tracing at a large scale. Nevertheless, I am not sure whether traitor tracing at large scale makes any sense.
  • The method is most probably robust against an oracle attack.
  • Some of the described methods were related to image recognition but could be ported to other tasks.
  • It is possible to embed several successive orthogonal watermarks.

A paper interesting to read as it is probably the beginning of a new field. ML/AI security will be key in the coming years.


Zhang, Jialong, Zhongshu Gu, Jiyong Jang, Hui Wu, Marc Ph. Stoecklin, Heqing Huang, and Ian Molloy. “Protecting Intellectual Property of Deep Neural Networks with Watermarking.” In Proceedings of the 2018 on Asia Conference on Computer and Communications Security, 159–172. ASIACCS ’18. New York, NY, USA: ACM, 2018.

Some notes on the Content Protection Summit 2015

These motes are personal and reflect the key points that raised my interest. They do not report the already known issues, already approved best practices and security guidelines.

The  conference was held on 7th December at Los Angeles. The audience was rather large for such event (more than 120 attendees) with representatives of content owners, service and technology providers and a few distributors. CPS is becoming the annual event in content protection. The event was as interesting as last year.

A special focus has been placed on cyber security rather than purely content protection.

Welcome remarks (ROSE M.)

The end of EU safe harbor is an issue.

CDSA: A focus on the right things at the right time (by ATKINSON R.)

A set of work streams for 2016 with nothing innovative. Some focus on training and education. A second focus on opportunity versus piracy.

IP security the creative perspective (by McNELIS B.)

An attack against YouTube that does not have in place a strong enough position against piracy. Google does not play the game despite it could (for instance, there is no porn on YouTube, proving the efficiency of curation). The difference between Apple and Google is the intent.

Creators do usually not want to bother about content protection. They want to communicate directly with consumers. The moderator explained that indie filmmakers are far more concerned as piracy may be more impacting their revenue stream. The middle class of creators is disappearing.

The BMG / Cox communication legal decision is a good promising sign.

Breakthrough in watermark (by OAKES G.)

NNSS (Nihil Nove Sub Sole, i.e., nothing new under the sun)

The move to digital pre-release screeners: DVD R.I.P. (panel with ANDERSON A., TANG E., PRIMACHENKO D.)


  • Nobody any more uses exclusively DVD at home, they use additional media. The user experience of DVD is bad (dixit Fox).
  • E-screener is more eco-friendly than DVD distribution.
  • Less liability due to no need to dispose of the physical support.
  • Higher quality is possible.
  • According to Fox, on-line screeners are intrinsically more secure than DVD screeners.


  • The challenge is the multiplicity of platforms to serve. Anthony pleads for 2FA.
  • Some guild members want to build a library.
  • Connectivity is still an issue for many members.

Suspicious behavior monitoring is a key security feature.

The global state of information security (by FRANK W.)

Feedback on the PcW annual survey of 40 questions.

  • Former employees are still the most cited sources. Third party related risk is rising.
  • Theft of employee and customer records raised this year.
  • 26% of increase of security budget over 2014.
  • ISO27001 is the most used framework. 94% of companies use a security framework.
  • Top Cyber threats: vulnerabilities, social engineering and zero-day vulnerabilities.
  • Data traversal becomes a visible issue with leaks via Dropbox, Google Drive…)

Would you rather be red and blue, or black and blue (by SLOSS J.)

A highlight on high-profile attacks. A plea for having an in-house red team (attack team)

He advocates the stance of assuming that you’re already penetrated. This requires:

  • War game exercises
  • Central security monitoring
  • Live site penetration test (not really new)

Secrets to build an incident response team (panel with RICKELTYON C., CATHCART H., SLOSS J.)

An Incident Response Team is now mandatory together with real-time continuous monitoring.

Personalize the risk by making personal what the consequences of a breach would be.

Hiring experts for a red team or IRT is tough.

Vulnerability scanning penetration testing (panel with EVERTS A., JOHNSON C., MEACHAM D., MONTECILLO M.)


Best practice for sending and receiving content (by MORAN T.)


  • Consumer grade cloud services: Dropbox, etc
  • Production. Media deal, signiant, mediafly, etc
    • Usually isolated system within a company
    • Owned by production rather than IT
  • Enterprise: Aspera
    • Owned by IT

Cooperation between IT and production staff is key.

Don’t tolerate shadow IT. Manage it

Monitor the progress of Network Function Virtual (NFV)and Software Defined Network (SDN) as they may be the next paradigms

Production in the cloud (panel with BUSSINGER B., DIEHL E., O’CONNOR M., PARKER C.)

CDSA reported about this panel at

Production security compliance (panel with CANNING J., CHANDRA A., PEARSON J., ZEZZA L.)

It is all about education. The most challenging targets are the creatives

New Regency tried on a production of a TV show to provide all creatives with the computer, tablet, and phone. They also allocated a full-time IT guy.

Dr Who’s leaked

Bad week for the BBC.   Last week, scripts of five episodes of next season of Dr Who leaked online.  The scripts were accessed from a Miami-based BBC worldwide server.  It seems that that they were publicly available (with a lot of material) and was indexed by Google.   A typical Google request provided access to this confidential material.

Unfortunately, other material was available.  A black & white unfinished watermark version of the first episode has also been put online.  The copy is visibly watermarked for a given recipient.   Drei Marc is a Brazilian company that provides subtitling and dubbing services.  Nevertheless, it seems that it comes from the same server.  It is not sure that other episodes may not surface in the coming days.  Broadcast of the first episode is planed on 23 August.


BBC asked its fans not to spoil the release.

We would like to make a plea to anyone who might have any of this material and spoilers associated with it not to share it with a wider audience so that everyone can enjoy the show as it should be seen when it launches.

"We know only too well that Doctor Who fans are the best in the world and we thank them for their help with this and their continued loyalty

Several lessons:

  • Secure your servers and be aware of the indexing robots.   No server should be put online without prior pen testing.
  • Encryption at rest should be mandatory for early content.  If ever the attacker access the video server, he will access an encrypted video without the decryption key.  Useless.
  • Forensic marking should only occur at delivery time.  If prepared and stored before release, it is useless.  It will not hold in front of a Court with good security expert.
  • TV series are the new Eldorado of the movie industry

World of Warcraft and watermarking

An old news, as it started in September.  On 8 September 2012, Sendatsu published on the ownedcore a detailed study of the use of watermark within Blizzard’s World of Warcraft (WoW).  According to him, it seems that WoW adds an “invisible” watermark to screenshots (at least with JPEG in lower quality).   A capture of a screenshot without texture repeatedly produces a pattern similar to this one.  wow-watermark

The watermark carries 88 bytes with the account ID, a time stamp and the IP address of the server.  Clearly, it does not carry any personal information.   It seems that this Digimarc based watermark was in use since 2007 (when screenshots were added).

The aim of this watermark seems obvious to me.  There are many illegal WoW servers in the field.  Of course, users playing WoW through these non-Blizzard servers do not pay the monthly subscription.  This means a loss of revenue for Blizzard.  Finding the IP address of such unauthorized servers is a good start to fight piracy back .

Strangely, nobody reported a similar case for other Blizzard MMORPGs such as Diablo III or StarCraft.  Is it because nobody looked at, it yet? Or because there is no such watermark (less pirate servers)?

Update (30-oct-12):  The allegation that it is a Digimarc solution seems wrong.  Thus, currently no clue about the solution provider.

Nano counterfeiting feature

The blue  morpho butterfly changes the color of iits wings through some special reflective structure.  The company nanotech security uses a “similar” trick for its NOtES (Nano Optic Technology for Enhanced Security).   Using nano holes smaller than the light wave, it creates a kind of light-amplification that generates a similar effect.


Thus, by embossing paper or plastic, it can create bright images through reflection.  The holes are about a few hundred nanometers.  How does it fit with security?   According to them, it could replace holograms used against counterfeiting (the kind of holograms that you find on microsoft official disks).  This technology seems to have some advantages:

  • It is extremely cost effective.  Once the master stamping build, it is just stamping the target, thus cheap and fast in production.
  • Easily identifiable by human
  • As it works infrared or UV, the pattern could be analyzed by machines using the right wave length (a kind of watermark)


The security relies on the difficulty for the counterfeiters to reproduce the stamping.  It seems that it relies mainly on a high barrier entry cost (class 1 clean room) and know how of the company to design the pattern and the stamping tool.   Clearly, it would require a funded organization to make it (as holograms today).   Nevertheless, I would be interested to see if it would be not possible to reverse engineer the pattern by careful examination through electronic microscope. Another question is how does it degrade with time?     

When will we have the first shiny bank notes?

“Securing Digital Video” is now available!

My book, “Securing Digital Video: Techniques for DRM and Content Protection” is now available on sale.   It can be found directly at Springer (about one week delay), from US amazon (2-4 weeks delay) and from French Amazon (available only in August).

This is the last step of a long process.  I hope that the reader will enjoy it and that it will be useful to the community.   More details on the book are available here.

I would be glad to hear your suggestions, appreciations (even negative ones), and answer any question.  For that, use preferably the address  I will always answer.