This latest Authenticate update introduces the new Reflections filter. Now, you confidently identify inconsistencies and prove when something isn’t quite right. We’ve also improved the Diffusion Model Deepfake filter, empowered Annotations, and more!
Dear friends, here we are with a shiny new update for Amped Authenticate! With this release, we’re enriching the Geometrical Analysis category with the new Reflections filter. Also, we’ve redesigned and updated the Diffusion Model Deepfake to ensure you’re provided with state-of-the-art deepfake detection capabilities! Keep reading to find out more!
See the New Features in Action
The New Reflections Filter
Creating consistent reflections when forging an image is a very challenging task. Luckily, this seems true even when the generation is done by the most recent AI generative models, such as Midjourney, DallE, and Stable Diffusion. We demonstrated this in our blog and in a scientific paper we recently presented.
AI Models Struggle with Reflections
Yes, diffusion models have achieved impressive accuracy and create visually convincing images. But they’re not designed to enforce physical constraints such as shadows, reflections, and so on. Indeed, a study we presented at the IEEE International Workshop on Information Forensics and Security revealed that commercial diffusion models have a very hard time getting reflections right.
For these reasons, we decided it was time to empower Amped Authenticate’s users with a dedicated filter.
How the Reflections Filter Works
Let’s clarify the filter scope first. The Reflections filter can be used to analyze images where you have a planar reflection. Translated to layman’s terms, that means mirrors, flat windows, or even a lake surface, provided there are no waves.
When objects are reflected in a mirror, we can model the situation as in the image below.
The image plane is where your digital image lies. In the 3D world, lines that connect each point to its reflection are parallel (the purple and red lines in the drawing above). Their projections onto the image plane (green lines), however, are no longer parallel: they intersect in the vanishing point (v).
The situation is very similar to our analysis of shadows. In that case, the lines connecting each cast shadow to their originating points converged to the light source even in the 3D world.
Therefore, the Reflections filter borrows the same technical approach as the Shadows filter. Notably, since it may be tricky to determine exactly which point in the reflection corresponds to a point in the object, we maintained the wedge-based selection for improved robustness.
Adding Reflections Constraints
If you’re familiar with the Shadows filter, using the Reflections filter will be very intuitive. You just need to click on a point of an object, then move to the corresponding reflected point and click twice to define the wedge that contains the reflected point, as illustrated below.
Please note that it is important that your first click is on the source object, not on the reflected one.
In our example, even after adding a few reflection constraints, the system remains feasible. This means that the purple wedges have a common intersection where the vanishing point could be located.
Using the Feasible Wedge
If you enable the “Feasible Wedge” checkbox, you’ll see a yellow wedge next to the mouse pointer that “previews” where the reflection wedge should be located based on the mouse position. This makes the search for inconsistencies much quicker. In fact, we notice that there may be an issue near the bottom eyelash: the yellow feasible wedge does not contain the eyelash in the reflection!
We jump on this opportunity and add a wedge for the eyelash. As expected, that makes the system unfeasible. This proves that the reflections are not consistent, and so the image should not be trusted.
By clicking on the Show Min. Unfeasible Set button, Authenticate will show you a minimal set of constraints that makes the system unfeasible.
Remember you can add constraints from multiple objects, provided they are all reflected by the same mirror, as in the example below. In our experiments, we found that multi-object consistency is particularly difficult to achieve for AI generative models (at the moment).
Improved Diffusion Model Deepfake Filter
Deepfakes’ impact on (dis)information is sensibly growing as AI-generated images become increasingly visually reliable. To help you combat the most recent form of deepfakes, created with diffusion models, we are improving how we present results in the Diffusion Model Deepfake filter.
After running the filter, the first line will now show you two possible outcomes:
- “Compatible with a known diffusion model”: this indicates that the AI-based classifier found the image to be compatible with one among Dall-E, Midjourney, and Stable Diffusion. You can check the scores in the remaining rows for the respective compatibility scores. Since diffusion models may share a part of the generation process, an AI-generated image may expose stronger compatibility with a different diffusion model than the one actually used to generate it.
- “Not compatible with a known diffusion model”: this indicates that the classifier found the image to be more compatible with the non-AI-generated images seen during the training phase. That doesn’t necessarily mean that the image is unaltered. It could have been altered with classical image editing software or partially edited with AI rather than being entirely generated.
Remember that this filter uses AI to detect AI. That means its output should be considered a clue to stimulate further analysis rather than a conclusive outcome. Also, remember that a compatibility score close to or equal to 1.00 does not mean you can be sure the image is AI-generated or vice versa. That’s just the classifier’s output. It doesn’t carry any “probabilistic” meaning. To understand why discussing probabilities can be problematic, especially in the context of deepfakes, watch our recent webinar below.
Relocating Missing Files
When transferring your Amped Authenticate project file to a different location or computer, you may accidentally move the associated image or video files to another location or rearrange them in the folder. When loading the project, Amped Authenticate will now inform you when a linked file is missing. You will have the option to either relocate the file or load the project, discarding the bookmarks connected to that file.
Remember that Authenticate also automatically checks the MD5 hash to ensure the file was not altered since you saved the project. This protects you from accidentally linking a file with the same name as the missing one.
This improvement applies to both the Image and Video modes.
New Blacken Effect in Annotate’s Hide
When annotating a bookmark, the Hide tool lets you redact sensitive information such as a subject’s face or a license plate. With this Authenticate update, we’re adding the “blacken” effect. This feature will turn your sensitive content into black pixels; you can’t go safer than this. 😉
To blacken an object, select the filter you want to bookmark from the Filters panel, add a bookmark from the Project panel and click on the pencil icon to open the Annotate view. Click on the Hide tool, choose “Blacken” from the Effect menu, and draw over the image.
Other Improvements
Updated libraries
We are updating several libraries at the core of Authenticate to ensure we keep software accuracy, compatibility, and performance at the state of the art.
- Core: updated the FFmpeg library to version 7.0.
- Core: updated the libheif library to version 1.18.2
- Core: updated the OpenCV library to version 4.9.0
GUI Improvements
We’ve replaced some textual buttons with visually intuitive icons and made some GUI elements resizable (e.g. the “View All QTs in Database” pop-up window, or the Shadows right panel). We’ve also improved the plotting system for the Video Mode so that decimal positions are shown only when necessary.
Bugfixes
We fixed some minor bugs, and once again, we thank all users who reported them to us!
- Image module: fixed a bug that caused side effects on the GUI when dragging bookmarks in the project panel.
- Image module: fixed a bug that caused a crash after typing in a value outside the spin control’s range.
- Image module: fixed a bug causing a bookmark with annotations to have an incorrect zoom level.
- Image module: fixed a bug causing the viewer to flicker with a previously opened image while switching between bookmarks.
- Image module: fixed a bug that kept the JPEG Ghosts Map filter stuck at 51% for some files.
- Image module: fixed a bug causing desynchronization of double viewer’s scrollbars.
- Video module: fixed a bug causing a misalignment between VPF and Macroblocks filters.
- Video module: fixed a bug in the Recent Projects menu not showing the full path for the first item.
- Video module: fixed a bug causing checkboxes to remain active after Macroblock filter deactivation with Show Input.
Don’t Delay – Update Today
If you have an active support plan, you can update straight away by going into the menu Help>Check for Updates Online 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.
