Detecting the New Wave of Hyper-Realistic Deepfake Videos

Real-time deepfake generation is getting scarily good. In the last couple of weeks our Threat Intelligence team has been tracking the emergence of some hyper-realistic deepfake full-body reenactment generators capable of creating videos that are convincing even to the trained eye. 

As you can see in the video, these novel models can create convincing full body avatars that can be puppeted by an operator and used to make standalone videos, or even injected real-time into a video stream using camera hijack techniques. 

These tools are already well adopted by cyber criminals to carry out various forms of fraud including romance scams, impersonation and identity theft, and employment fraud or corporate subterfuge. 

Other threat actors may also use such convincing deepfake videos to spread disinformation or influence public and political sentiment, especially when masquerading as public figures. 

New wave hyper-realistic deepfakes enable more sophisticated attacks

But it’s not just the hyper-realistic quality of the media that’s concerning, it’s the ease with which these deepfakes can be created. Even the consumer-friendly gen-AI tools intended for entertainment and self-labelled as such, can be abused by malicious actors for more sinister purposes by removing the embedded watermarks. 

There’s also considerable sophistication in attack vectors. Adversaries can easily inject these face- and body-reenactment video streams into real-time KYC (Know Your Customer) systems and video conferencing platforms. So, threat analysts need to be able to detect deepfakes in real-time and at scale.  

Detecting novel gen-AI deepfake models

Fortunately, Sensity AI’s Deepfake Detection engine is already detecting these novel models and flagging this media as suspicious. Because deepfake detection is an arms race, detection improves with training on datasets generated by each model, which is why you see varying levels of confidence. 

In the first example you can see the Sensity AI Deepfake Detection platform detects the media as fake, even though it has low confidence in identifying the specific generator, due to its novel nature. Our analysts believe it is the Kling video AI model generator, developed by China’s Kuaishou Technology, that was used. 

Early detection indicators like this are testament to Sensity AI’s strategy for robust detection, in which we have equipped our solution with multiple layers of analysis, looking at elements like pixels, voice, face manipulation, bytecode analysis, C2PA certificates, watermarks and more, leveraging deep neural networks and digital forensics. 

The goal is that if one layer proves inconclusive or provides a low confidence analysis, another layer will trigger with a more definitive result, providing a more robust overall assessment. Ultimately, this is what makes our solution capable of meeting forensic standards in courts of law.

So, what you can see in the first video is that the AI-generated content check has analyzed the entire video frame-by-frame for any synthesis signals and come up with an average score of 52.1% across all frames. However, our explainability features like heat maps, which show pixel-level noise most interesting to the model, and segment-based analysis, which highlight likely AI-generated objects, provide strong forensic signals even when the overall confidence score is low.

For example, the heatmap has identified the cat-ear headphones as suspicious, and on close inspection with the human eye you can see they appear fabricated. But it’s the segment-based analysis that flags the avatar and the environment as AI-generated. 

In the second example, the confidence is much higher at 87.8%, clearly detecting evidence of face manipulation, despite the model being novel, while also exposing pixel-level noise as further evidence of synthesis.