Category: Blog Post

Contents

• The Danger of Deepfakes
• A Look Into Our Model
• Putting It All Together: Example Input and Response
• Final Thoughts

The Danger of Deepfakes

When generative AI models first gained popularity in the late 2010s, they brought with them the ability to create deepfakes. Deepfakes are synthetic media, typically video, in which one person’s likeness is replaced by another’s using deep learning. They are powerful tools for fraud and misinformation, allowing for the creation of synthetic videos of political leaders and letting scammers easily take on new identities.

The primary use, though, of deepfake technology is the fabrication of nonconsensual pornography. The term “deepfake” itself was coined in 2017 by a Reddit user of the same name who made fake pornographic videos featuring popular female celebrities. In 2019, the company Sensity AI catalogued deepfakes across the web and reported that a whopping 96% of them were pornographic, all of which were of women. In the years since, more of this sort of deepfake pornography has become readily available online, with countless forums and even entire porn sites dedicated to it. The targets of this are not just celebrities. They are also everyday women superimposed into adult content by request—on-demand revenge porn for anyone with an internet connection.

Many sites have banned deepfakes entirely, since they are far more often used for harm than for good. At Hive, we’re committed to providing API-accessible solutions for challenging moderation problems like this one. We’ve built our new Deepfake Detection API to empower enterprise customers to easily identify and moderate deepfake content hosted on their platforms.

This blog post explains how our model identifies deepfakes and the new API that makes this functionality accessible.

A Look Into Our Model

Hive’s Deepfake Detection model is essentially a version of our Demographic API that is optimized to identify deepfakes as opposed to demographic attributes. When a query is submitted, this visual detection model locates any faces present in the input. It then performs an additional classification step that determines whether or not each detected face is a deepfake. In its response, it provides a bounding-box location and classification (with confidence scores) for each face.

While the face detection aspect of this process is the same as the one used for our industry-leading Demographic API, the classification step was fine-tuned for deepfake identification by training on a vast repository of synthetic and real video data. Many of these examples were pulled from genres commonly associated with deepfakes, such as pornography, celebrity interviews, and movie clips. We also included other types of examples in order to create a classifier that identifies deepfakes across many different content genres.

Putting It All Together: Example Input and Response

With only one head, the response of our Deepfake Detection model is easily interpretable. When an image or video query is submitted, it is first split into frames. Each frame is then analyzed by our visual detection model in order to find any faces present in the image. Every face then receives a deepfake classification — either yes_deepfake or no_deepfake. Confidence scores for these classifications range from 0.0 to 1.0, with a higher score indicating higher confidence in the model’s results.

Here we see the deepfaked image and, to its left, the two original images used to create it. This input image doesn’t appear to be fake at first glance, especially when the image is displayed at a small size. Even with a close examination, a human reviewer could fail to realize that it is actually a deepfake. As the example illustrates, the model correctly identifies this realistic deepfake with a high confidence score of more than 0.99. Since there is only one face present in this image, we see one corresponding “bounding poly” in the response. This “bounding poly” contains all model response information for that face. Vertices and dimensions are also provided, though those fields are truncated here for clarity.

Because deepfakes like this one can be very convincing, they are difficult to moderate with manual flagging alone. Automating this task is not only ideal to accelerate moderation processes, but also to spot realistic deepfakes that human reviewers might miss.

Digital platforms, particularly those that host NSFW media, can integrate this Deepfake Detection API into their workflows by automatically screening all content as it is posted. Video communication platforms and applications that use any kind of visual identity verification can also utilize our model to counter deepfake fraud.

Final Thoughts

Hive’s Deepfake Detection API joins our recently released AI-Generated Media Recognition API in the aim to expand content-moderation to keep up with the fast-growing domain of generative AI. Moving forward, we plan to continually update both models so as to best keep up with new generative techniques, popular content genres, and emerging customer needs.

The recent popularity of diffusion models like Stable Diffusion, Midjourney, and DALL-E 2 has brought deepfakes back into the spotlight and sparked conversation on whether these newer generative techniques can be used to develop brand-new ways of making them. Whether or not this happens, deepfakes aren’t going away any time soon and are only growing in number, popularity, and quality. Identifying and removing them across online platforms is crucial to limit the fraud, misinformation, and digital sexual abuse that they enable.

If you’d like to learn more about our Deepfake Detection API and other solutions we’re building, please feel free to reach out to sales@thehive.ai or contact us here.

Try Our Demo

To try our AI-Generated Image Detection model out for yourself, check out our demo.

Contents

• A New Need for Content Moderation
• Using AI to Identify AI: Building Our Classifier
• How it Works: An Example Input and Response
• Final Thoughts and Future Directions

A New Need for Content Moderation

In the past few months, AI-generated art has experienced rapid growth in both popularity and accessibility. Engines like DALL-E, Midjourney, and Stable Diffusion have spurred an influx of AI-generated artworks across online platforms, prompting an intense debate around their legality, artistic value, and potential for enabling the propagation of deepfake-like content. As a result, certain digital platforms such as Getty Images, InkBlot Art, Fur Affinity, and Newgrounds have announced bans on AI-generated content entirely, with more to likely follow in the coming weeks and months.

Platforms are enacting these bans for a variety of reasons. Online communities built for artists to share their artwork such as Newgrounds, Fur Affinity, and Purpleport stated they put their AI artwork ban in place in order to keep their sites focused exclusively on human-created art. Other platforms have taken action against AI-generated artwork due to copyright concerns. Image synthesis models often include copyrighted images in their training data, which consist of massive amounts of photos and artwork scraped from across the web, typically without any artists’ consent. It is an open question whether this type of scraping and the resulting AI-generated artwork amount to copyright violations — particularly in the case of commercial use — and platforms like Getty and InkBlot Art don’t want to take that risk.

As part of Hive’s commitment to providing enterprise customers with API-accessible solutions to moderation problems, we have created a classification model made specifically to assist digital platforms in enacting these bans. Our AI-Generated Media Recognition API is built with the same type of robust classification model as our industry-leading visual moderation products, and it enables enterprise customers to moderate AI-generated artwork without relying on users to flag images manually.

This post explains how our model works and the new API that makes this functionality accessible.

Using AI to Identify AI: Building Our Classifier

Hive’s AI-Generated Media Recognition model is optimized for use with the kind of media generated by popular AI generative engines such as DALL-E, Midjourney, and Stable Diffusion. It was trained on a large dataset comprising millions of artificially generated images and human-created images such as photographs, digital and traditional art, and memes sourced from across the web.

The resulting model is able to identify AI-created images among many different types and styles of artwork, even correctly identifying AI artwork that could be misidentified by manual flagging. Our model returns not only whether or not a given image is AI-generated, but also the likely source engine it was generated from. Each classification is accompanied by a confidence score that ranges from 0.0 to 1.0, allowing customers to set a confidence threshold to guide their moderation.

How it Works: An Example Input and Response

When receiving an input image, our AI-Generated Media Recognition model returns classifications under two separate heads. The first provides a binary classification as to whether or not the image is AI-generated. The second, which is only relevant when the image is classified as an AI-made image, identifies the source of that artificial image from among the most popular generation engines that are currently in use.

To get a sense of the capabilities of our AI-Generated Media Recognition model, here’s a look at an example classification:

This input image was created with the AI model Midjourney, though it is so realistic that it may be missed by manual flagging. As shown in the response above, our model correctly classifies this image as AI-generated with a high confidence score of 0.968. The model also correctly identifies the source of the image, with a similarly high confidence score. Other sources like DALL-E are also returned along with their respective confidence scores, and the scores under each of the two model heads sum to 1.

Platforms that host artwork of any kind can integrate this AI-Generated Media Recognition API into their workflows by automatically screening all content as it is being posted. This method of moderating AI artwork works far more quickly than manual flagging and can catch realistic artificial artworks that even human reviewers might miss.

Final Thoughts and Future Directions

Digital platforms are now being flooded with AI-generated content, and that influx will only increase as these generative models continue to grow and spread. On top of this, creating this kind of artwork is fast and easy to access online, which enables large quantities of it to be produced quickly. Moderating artificially created artworks is crucial for many sites to maintain their platform’s mission and protect themselves and their customers from potential legal issues further down the line.

We created our AI-Generated Media Recognition API to solve this problem, but our model will need to continue to evolve along with image generation models as existing ones improve and new ones are released. We plan on adding new generative engines to our sources as well as continually updating our model to keep up with the current capabilities of these models. Since some newer generative models can create video in addition to still images, we are working to add support for video formats within our API in order to best prevent all types of AI-generated artwork from dominating online communities where they are unwelcome.

If you’d like to learn more about this and other solutions we’re building, please feel free to reach out to sales@thehive.ai or contact us here.

Contents

• New Visual Moderation Classes for Greater Content Understanding
  • Undressed
  • Gambling
  • Confederate Symbolism
• Improvements to Established Visual Moderation Classes
• Final Thoughts

Hive’s visual classifier is a cornerstone of our content moderation suite. Our visual moderation API has consistently been the best solution on the market for moderating key types of image-based content, and some of the world’s largest content platforms continue to trust Hive’s Visual Moderation model for effective automated enforcement on NSFW images, violence, hate, and more. 

As content moderation needs have evolved and grown, our visual classifier has also expanded to include 75 moderation classes across 31 different model heads. This is usually an iterative process – as our partners continue to send high volumes of content for analysis, we uncover ways to refine our classification schemes and identify new, useful types of content.

Recently, we’ve worked on broadening our visual model by defining new classes with input from our customers. And today, we’re shipping the general release of our latest visual moderation model, including three new classes to bolster our existing model capabilities:

• Undressed to target racier suggestive images that may not be explicit enough to label NSFW
• Gambling to capture betting in casinos or on games and sporting events
• Confederate to capture imagery of the Confederate flag and graphics based on its design

All Hive customers can now upgrade to our new model to access predictions in these new classes at no additional cost. In this post, we’ll take a closer look at how these classes can be used and our process behind creating them.

New Visual Moderation Classes For Greater Content Understanding

Deep learning classifiers are most effective when given training data that illustrates a clear definition of what does and does not belong in the class. For this release, we used our distributed data labeling workforce – with over 5 million contributors – to efficiently source instructive labels on millions of training images relevant to our class definitions. 

Below, we’ll take a closer look at some visual examples to illustrate our ground truth definitions for each new class. 

Undressed

In previous versions, Hive’s visual classifier separated adult content into two umbrella classes: “NSFW,” which includes nudity and other explicit sexual classes, and “Suggestive,” which captures milder classes that might still be considered inappropriate. 

Our “Suggestive” class is a bit broad by design, and some customers have expressed interest in a simple way to identify the racier cases without also flagging more benign images (e.g., swimwear in beach photos). So, for this release, we trained a new class to refine this distinction: undressed. 

We wanted this class to capture images where a subject is clearly nude, even if their privates aren’t visible due to their pose, are temporarily covered by their hands or an object, or are occluded by digital overlays like emojis, scribbles, or shapes. To construct our training set, we added new annotations to existing training images for our NSFW and Suggestive classes and sourced additional targeted examples. Overall, this gave us a labeled set of 2.6M images to teach this ground truth to our new classifier. 

Here’s a mild example to help illustrate the difference between our undressed and NSFW definitions (you can find a full definition for undressed and other relevant classes in our documentation): 

The first image showing explicit nudity is classified as both undressed and NSFW with maximum confidence. When we add a simple overlay over relevant parts of the image, however, the NSFW score drops far below threshold confidence while the undressed score remains very high. 

Platforms can use undressed to flag both nudity and more obviously suggestive images in a single class. For content policies where milder images are allowed but undressed-type images are not, we expect this class to significantly reduce any need for human moderator review to enforce this distinction. 

Gambling

Gambling was another type of content that frequently came up in customer feedback. This was a new undertaking for Hive, and building our ground truth and training set for this class was an interesting exercise in definitions and evaluating context in images. 

Technically, gambling involves a wager staked on an uncertain outcome with the intent of winning a prize. For practical purposes, though, we decided to consider evidence of betting as the key factor. Certain behavior – like playing a slot machine or buying a lottery ticket – is always gambling since it requires a bet. But cards, dice, and competitive games don’t necessarily involve betting. We found the most accurate approach to be requiring visible money, chips or other tokens in these cases in order to flag an image as gambling. Similarly, we don’t consider photos at races or sporting events to be gambling unless receipts from a betting exchange or website are also shown.  

To train our new class on this ground truth definition, we sourced and labeled a custom set of over 1.1M images. The new visual classifier can now distinguish between gambling activity and similar non-gambling behavior, even if the images are visually similar:

For more detailed information, you can see a full description of our gambling class here. Platforms that wish to moderate or identify gambling can access predictions from this model head by default after upgrading to this model release. 

Confederate Symbolism

Separately, many of our customers also expressed interest in more complete monitoring of visual hate and white nationalism, especially Confederate symbolism. For this release, we sourced and labeled over 1M images to train a new class for identifying imagery of the commonly used version of the Confederate flag. 

In addition to identifying photos of the flag itself, this new model head will also capture the Confederate “stars and bars” shown in graphics, tattoos, clothing, and the like. We also trained the model to ignore visually similar flags and historical variants that are not easily recognizable:

Along with our other hate classes, customers can now use predictions from our Confederate class to keep their online environments safe.

Improvements to Established Visual Moderation Classes

Beyond these new classes, we also focused on improving the model’s understanding around niche edge cases in existing model heads. For example, we leveraged active learning and additional training examples to address biases we occasionally found in our NSFW and Gun classifiers. This corrected some interesting biases where the model sometimes incorrectly identified studio microphones as guns, or mistook acne creams for other, less safe-for-work liquids. 

Final Thoughts

This release delivers our most comprehensive and capable Visual Moderation model yet to help platforms develop proactive, cost-effective protection for their online communities. As moderation needs become more sophisticated, we’ll continue to incorporate feedback from our partners and refine our content moderation models to keep up. Stay tuned for our next release with additional classes and improvements later this year.

If you have any questions about this release, please get in touch at support@thehive.ai or api@thehive.ai. You can also find a video tutorial for upgrading to the latest model configuration here. For more information on Visual Moderation more generally, feel free to reach out to sales@thehive.ai or check out our documentation. 

Contents

• Content Moderation Challenges in the Aftermath of Buffalo
• Conventional Moderation and Copy Detection Approaches
• Deep Learning to Advance Video Moderation and Contain Viral Content
• How it works: Deep Learning Models in Automated Content Moderation Systems
• Final Thoughts: How Hive Can Help

Content Moderation Challenges in the Aftermath of Buffalo

The racially-motivated shooting in a Buffalo supermarket – live streamed by the perpetrator and shared across social media – is tragic on many levels.  Above all else, lives were lost and families are forever broken as a result of this horrific attack.  Making matters worse, copies of the violent recording are spreading on major social platforms, amplifying extremist messages and providing a blueprint for future attacks.

Unfortunately, this is not a new problem: extremist videos and other graphic content have been widely shared for shock value in the past, with little regard for the negative impacts. And bad actors are more sophisticated than ever, uploading altered or manipulated versions to thwart moderation systems.

As the world grapples with broader questions of racism and violence, we’ve been working with our partners behind the scenes to help control the spread of this and other harmful video content in their online communities.  This post covers the concerns these partners have raised with legacy moderation approaches, and how newer technology can be more effective in keeping communities safe. 

Conventional Moderation and Copy Detection Approaches

Historically, platforms relied on a combination of user reporting and human moderation to identify and react to harmful content. Once the flagged content reaches a human moderator, enforcement is usually quick and highly accurate. 

But this approach does not scale for platforms with millions (or billions) of users.  It can take hours to identify and act on an issue, especially in the aftermath of a major news event when post activity is highest.  And it isn’t always the case that users will catch bad content quickly: when the Christchurch massacre was live streamed in 2019, it was not reported until 12 minutes after the stream ended, allowing the full video to spread widely across the web.

More recently, platforms have found success using cryptographic hashes of the original video to automatically compare against newly posted videos.  These filters can quickly and proactively screen high volumes of content, but are generally limited to detecting copies of the same video. Hashing checks often miss content if there are changes to file formats, resolutions, and codecs. And even the most advanced “perceptual” hashing comparisons – which preprocess image data in order to consider more abstract features – can be defeated by adversarial augmentations.  

Deep Learning To Advance Video Moderation and Contain Viral Content

Deep learning models can close the moderation capability gap for platforms in multiple ways. 

First, visual classifier models can proactively monitor live or prerecorded video for indicators of violence.  These model predictions enable platforms to shut down or remove content in real-time, preventing the publishing and distribution of videos that break policies in the first place.  The visual classifiers can look for combinations of factors, such as someone holding a gun, bodily injury, blood, and other object or scene information to create automated and nuanced enforcement mechanisms. Specialized training techniques can also accurately teach visual classifiers to identify the difference ​​between real violence and photorealistic violence depicted in video games, so that something like a first-person shooter game walkthrough is not mistaken for an real violent event.

In addition to screening using visual classifiers, platforms can harness new types of similarity models to stop reposts of videos confirmed to be harmful, even if those videos are adversarially altered or manipulated. If modified versions somehow bypass visual classification filters, these models can catch these videos based on visual similarity to the original version.   

In these cases, self-supervised training techniques expose the models to a range of image augmentation and manipulation methods, enabling them to accurately assess human perceptual similarity between image-based content. These visual similarity models can detect duplicates and close copies of the original image or video, including more heavily modified versions that would otherwise go undetected by hashing comparisons.

Unlike visual classifiers, these models do not look for specific visual subject matter in their analysis.  Instead, they quantify visual similarity on a spectrum based on overlap between abstract structural features. This means there’s no need to produce training data to optimize the model for every possible scenario or type of harmful content; detecting copies and modified versions of known content simply requires that the model accurately assess whether images or video come from the same source.

How it works: Deep Learning Models in Automated Content Moderation Systems

Using predictions from these deep learning models as a real-time signal offers a powerful way to proactively screen video content at scale. These model results can inform automated enforcement decisions or triage potentially harmful videos for human review. 

Advanced visual classification models can accurately distinguish between real and photorealistic animated weapons. Here are results from video frames containing both animated and real guns. 

To flag real graphic violence, automated moderation logic could combine confidence scores in actively held weapons, blood, and/or corpse classes but exclude more benign images like these examples. 

As a second line of defense, platforms need to be able to detect reposts or modified versions of known harmful videos from spreading.  To do this, platforms use predictions from pre-trained visual similarity models in the same way they use hash comparisons today. With an original version stored as a reference, automated moderation systems can perform a frame-wise comparison with any newly posted videos, flagging or removing new content that scores above a certain similarity threshold.  

In these examples, visual similarity models accurately predict that frame(s) in the query video are derived from the original reference, even under heavy augmentation. By screening new uploads against video content known to be graphic, violent, or otherwise harmful, these moderation systems can replace incomplete tools like hashing and audio comparison to more comprehensively solve the harmful content detection problem.

Final Thoughts: How Hive Can Help

No amount of technology can undo the harm caused by violent extremism in Buffalo or elsewhere.  We can, however, use new technology to mitigate the immediate and future harms of allowing hate-based violence to be spread in our online communities. 

Hive is proud to support the world’s largest and most diverse platforms in fulfilling their obligation to keep online communities safe, vibrant, and hopeful. We will continue to contribute towards state-of-the-art moderation solutions, and can answer questions or offer guidance to Trust & Safety teams who share our mission at support@thehive.ai.