Authenticate Update 40960: Introducing Video Deepfake Detection, a New Filter for Detecting the Adobe Watermark, Faster Image Deepfake Detection, and more!

Reading time: 9 min

With the improvements introduced in this Authenticate update, you can screen video frames for diffusion-model deepfake traces, examine Adobe RAW conversion watermark patterns for possible tampering, and run image deepfake analysis faster with GPU support. The update also adds Viewer controls, PRNU support for flipped images, and refinements for clearer authentication results and reporting.

Dear friends, welcome to another fantastic update of Amped Authenticate! There’s a lot to say:

• We’re excited to share that deepfake detection enters the Video Mode with a dedicated Diffusion Model Deepfake filter;
• We’ve added a new filter to the Image Mode called Raw Conversion Watermark;
• We improved the performance and caching capabilities of deepfake detection 
• We’ve empowered the Viewer to let the user choose if the Reference and Evidence image sizes should be synced
• … and more!

Let’s dive into the details!

Key Takeaways

• Amped Authenticate now brings diffusion-model deepfake detection to Video Mode. You can analyze a full video or a selected frame range and review frame-level results through the Viewer, Plot, and Table panels.
• Video deepfake detection is designed to support screening, not replace forensic interpretation. Use the results to focus your review, then look for explainable traces such as inconsistencies in shadows, perspective, compression, or other visual and technical indicators.
• The new Raw Conversion Watermark filter helps examine images developed from RAW files with Adobe products. When the watermark is detected globally, the filter can show where the pattern is weaker or missing, helping you spot areas that may deserve closer attention.
• Deepfake analysis in Image Mode is faster when a compatible NVIDIA GPU is available. This is especially useful when you need to analyze all images in an evidence folder.
• The update also improves day-to-day authentication review. You get more control over Reference and Evidence image scaling, PRNU support for flipped images, clearer filter behavior in Video Mode, improved legends, and several fixes that make results easier to read and report.

See the New Features in Action

New Diffusion Model Deepfake Filter in the Video Mode

Authenticate users are already familiar with the Diffusion Model Deepfake filter: it’s been there in the Image Mode for years now, and has been continuously updated. From time to time, users asked whether they could use it to analyze individual video frames, and we always replied the filter wasn’t trained for that task, so we wouldn’t recommend it.

Now, the problem isn’t there anymore! In the Video Mode, there’s a sibling Deepfake Detection category featuring the Diffusion Model Deepfake filter for videos!

It shares the very same backbone and functioning as the image mode version. However, it has been trained on thousands of samples to detect traces of diffusion model generation at the individual frame level. Among the deepfake generation systems we’ve used to train it, we have: Grok, SORA 2, Veo 3.1, Imagine May 2026, Happy Horse 1.0, Kling 3.0, LUMA Ray 2, and many others.

Let’s have a look at how it works.

After loading a video, go to the Deepfake Detection category and click on the Diffusion Model Deepfake filter. The next thing to do is to decide whether you want to analyze the whole video, or just a range of frames. If you’re interested in a range, use the range selection controls in the player panel to set the beginning and end. Once ready, simply click on the Run Analysis button and wait for the process to finish.

This filter makes the most of Authenticate Video architecture. Its results are displayed in the Viewer as an overlay, consistent with how it works in the Image Mode (you can turn the overlay on and off from the filter settings, without triggering any recomputation). Then you have results shown in the Plot panel, so you can quickly see if all the analyzed frames seem to come from a deepfake generator, or only some of them. Finally, you have a summary of the results in the Table viewer. It shows the detection model used, the number of analyzed frames, how many of them are above the threshold, and the median and average confidence of frames being compatible with an AI model.

As you can see in our example above, the plot shows a sharp transition from very low scores to much higher ones, which are well above the threshold. Let’s move forward in the video and see who Marco was speaking to.

Ah! Although it’d be lovely to have a chat with Arnold, this video was actually obtained by collating an authentic part taken from the podcast about Deepfake Forensics featuring Marco Fontani and Martino Jerian, and a part where Martino’s face was replaced with Arnold Schwarzenegger’s face.

Since the Diffusion Model Deepfake filter operates at the frame level, the plot quickly helps us identify which frames are more likely to come from an AI model and which are less likely.

In line with how results are displayed by the similarly named filter in the Image Mode, you will notice a gray or red border surrounding the frame in the Viewer. A red border indicates that the score for “Compatible with known AI model” exceeds the threshold, while a gray border means it falls below the threshold.

The Importance of Using a GPU

The Diffusion Model Deepfake filter is based on a deep learning algorithm, so it is highly recommended to use a computer with a reasonably recent NVIDIA GPU to run it. If a compatible GPU is not found, Authenticate will fall back to using the CPU, with considerably longer processing times.

We’ve also improved the Image Mode version of the same filter to use the GPU when available. This will be especially beneficial when using the “Analyze all images in evidence folder” feature.

The Role of AI Detection in Fighting Deepfakes

Please always remember that, just like its Image Mode counterpart, this Diffusion Model Deepfake filter in Video Mode is a machine learning-powered detector. As we explained in the podcast, its role is not to be the sole indicator for labeling a video as a deepfake. Nonetheless, this kind of analysis is invaluable to perform an initial screening and focus the attention on looking for more explainable results such as inconsistencies in shadows, perspectives, etc.

New Raw Conversion Watermark filter

You may remember that a few years ago, we collaborated with the University of Florence (Italy) to publish a very important paper. The study revealed that many digital cameras, mostly smartphones but also some DSLRs, can produce false positives in PRNU analysis.

That paper triggered a lot of research worldwide, and while investigating the phenomenon related to DSLR cameras, researchers from the University of Lille have discovered a very interesting fact. When Adobe products like Lightroom and Premiere are used to convert a RAW 16-bit image into an 8-bit image (which is the standard workflow to obtain “universally compatible images”), a patch-wise invisible watermark is subtly embedded into the image (!!!).

They have presented their findings in a scientific paper titled “Detection of the Adobe Pattern”, accessible here:

The watermark is 128 by 128 pixels and is applied patchwise to cover the whole image. Experiments in the mentioned paper show that it’s very robust to JPEG compression

Let’s imagine this scenario: a photographer takes a picture with a professional camera in raw format. They then use Photoshop or Lightroom to develop the image and create a JPEG (or TIFF, or whatever else format) file. During this process, the 128-by-128 pattern is injected into the image. Someone then opens up the image in whatever editor, tampers with it, and saves it again. The pattern will be destroyed in the manipulated area, while it will remain elsewhere.

Now, if we look for the pattern patch-wise throughout the image, we can build a forgery localization map: blocks where the pattern is detected are likely authentic, while those where it’s missing are suspicious. The whole process is summarized in the (admittedly AI-generated 😉) infographic below.

For example, this image from the RAISE dataset had been manipulated to remove a person who was walking in the scene. The RAW Conversion Watermark filter clearly highlights in red the region where the watermark is not found.

Of course, it makes sense to search locally for the watermark when it’s known to be there globally; otherwise, you’d only get a mostly reddish, meaningless map. To facilitate your work, the filter works in two phases:

• First, it checks if the watermark is detected globally. If it’s not, a black image will be output, similarly to how the JPEG Dimples Map filter does.
• If the watermark is detected globally, the block-wise detection map will be shown, where greener blocks indicate the watermark has been detected more strongly.

We downloaded from the FloreView dataset 300 random images, which are known to be captured by various smartphone models: the filter mistakenly detected the watermark as being globally present in only 1% of the images. We also downloaded 300 images from the Unsplash Lite dataset, where images are typically captured with DSLR cameras. In this case, the filter globally detected the watermark in roughly 30% of the images, which is in line with numbers reported by researchers in their study.

Before concluding this section, a couple of remarks are needed:

• Since it is harder to detect the watermark in white- or black-saturated areas, we recommend activating the “Show saturation” feature, which highlights black-saturated regions in blue and white-saturated regions in yellow.
• While the watermark can survive strong JPEG compression, the detection will fail if the image undergoes any geometric transformation, such as downscaling, flipping, or rotation. We’ll look into increasing the detector’s robustness in the future.

Other Improvements

Ability to Turn Reference Image Scaling On and Off

This has been requested a few times by our users, so here we go! There’s a new button in the Image Mode that allows the toggle of the pixel resolution synchronization between the reference image and the evidence image. By default, the synchronization is now off!

When the reference image is downscaled, the Area interpolation algorithm is used to reduce aliasing effects.

Keep the feature disabled if you want to compare side-by-side images with different aspect ratios. Otherwise, the reference image will be deformed to match the size of the evidence image.

Face GAN Deepfake: Adjustable Font Size

We’ve added a dedicated slider to adjust the font size of the text labels in the filter parameters.

PRNU Source Identification: Added Support for Flipped Images

In both the Image and Video Modes, the PRNU Identification filter now supports detection for images that have been flipped.

Other Improvements and Bug Fixes

We’ve made other improvements to the software, including:

• Image Mode: the Diffusion Model Deepfake filter now uses the GPU (when available and compatible) to speed up processing.
• Video Mode: we’ve improved the filter tree behavior and added filter descriptions.
• Macroblocks and VPF: we’ve improved the legend.
• When creating a new project, the Evidence and Reference files are now automatically cleared.
• Macroblocks and Motion Vectors: improved the automated analysis that prevents the filters from being used on unsupported codecs.
• We reduced the number of decimals displayed in several filters to obtain a more readable output.
• We fixed a bug affecting the “Show Minimum Unfeasible Set” feature of the geometric filters.
• We fixed a bug that caused the “Reprocess filter Configuration” command not to work for the Face GAN Deepfake filter

Don’t Delay – Update Today

The new features make Authenticate more powerful than ever.

If you have an active support plan, you can update straight away by going into the menu About > Check for Updates within Amped Authenticate. If you need to renew your SMS plan, please contact us or one of our authorized distributors. And remember that you can always manage your license and requests from the Amped Support Portal.

---

FAQ – Amped Authenticate Update Authenticate Update 40960