Back in 2017, Sensity AI co-founder Giorgio Patrini was a postdoctoral researcher working on deep generative models with Max Welling at the University of Amsterdam.
This pioneering research focused on image generation with neural nets, which at the time amounted to creating tiny profile photos with GANs (Generative Adversarial Networks) and VAEs (Variational Autoencoders). But this work, and research like it, set the path for generative models that would soon power every use case for content creation across images, videos, and audio, and put the power for what would become ‘deepfake’ generation into the hands of everyone.
At that time, the tech community at large wasn’t reflecting on the likely consequences of this development: that within a decade, digital communication would fundamentally break down once content could be easily fabricated. ‘Reality’ would no longer be objective. What you see and hear could no longer be trusted.
Patrini realized that putting generative-AI tools into the hands of malicious actors would inevitably give them new ways to coerce, defraud, and attack others. A counter movement needed to happen.
The birth of deepfake detection
In March 2018, Patrini became possibly the first person ever to train a deepfake detector, a development which showed promising results. Working on the assumption that advances in technology would rapidly increase the realism of AI-generated media, making it exponentially harder for humans to identify synthetic images and audio, Patrini became convinced that there would soon be a need for forensically robust deepfake detection technology that could hold up explainability in a court of law.
Getting ahead of the curve, Patrini joined forces with Francesco Cavalli, a threat intelligence specialist researching the phenomenon of ‘fake news’ and its social effects on internet communities and politics, and late in 2018, Sensity AI was born as the world’s first deepfake detection company.
Mission
Sensity AI exists to protect states, organizations, businesses, and individuals from malicious deepfakes.
Vision
A world where all digital media can be authenticated in real-time, to prove conclusively whether it is manipulated by generative-AI or not.
Why?
We believe AI-generated media and deepfakes are a societal problem of the same scope as something like climate change – inevitable and world-changing. The democratization of AI-generated media and deepfakes is re-defining the meaning of trust in what we see and hear, anytime we do not experience it first hand.
Sensity AI is a response to synthetic reality
We see malicious deepfakes and similar synthetic media technologies as a new wave of cybersecurity threats, with the potential of affecting every digital audiovisual communication channel used by every human.
Just as antivirus software has become the ubiquitous solution for scanning files obtained outside of trusted sources, we envision a similar frictionless verification approach for audiovisual media. What we are defending is not software that can be infected and manipulated by malware, but human sentiment, opinions, and actions, infected and manipulated by fake videos, impersonation, influence, and sophisticated cyberfrauds.
Crucially however—and this is a key differentiator in the industry of ‘deepfake detection’—the solution we are building is forensically robust. We don’t believe it is enough to be able to say “this media is synthetic” or not, our analysis has to be conclusive and explainable enough to meet the test of evidence in a court of law. Otherwise it would not provide any real security to the victims of complex AI-generated cyberfraud, it would just be a checkbox solution. Our real purpose is galvanized around meaningful impact.
Sensity AI’s deepfake detection technology works by detecting the ‘digital fingerprints’ left behind on synthetically manipulated or generated photos, videos, and audio. We do this by training a deep learning network on large sets of known deepfake data, where it learns to identify these fingerprints at the pixel level, or by analyzing subtle acoustic artifacts. As deepfake generation technology continues to improve and the visual and aural clues of manipulation inevitably become undetectable by humans, this machine-level analysis will become essential for reliably detecting deepfakes.
But this also needs to happen at scale. One striking element of the current state of deepfake creation technology compared to just 12 months ago, is just how little technical knowledge an individual requires to create synthetic media. Millions of innocuous and malicious deepfakes are now created every day, and this has created an unworkable caseload for forensic investigators and law enforcement specialists tasked with enforcing justice for victims of child sexual abuse, harassment, coercion, and fraud.
We’re not only writing the playbook for accurate detection of malicious deepfake media, we’re also helping authorities evolve the forensic process to keep pace in a fast-moving market and continue delivering results.
Our advantage is that we are a deep learning company first, and that ensures we can keep pace with a rapidly evolving threat vector. But while we are developing a technological solution, we also have strong knowledge of the greater socio-ethical impact of deepfakes, and how their impact will be felt in numerous different ways.
How is Deepfake Detection Changing Forensic Analysis?