Indust.ai’s journey into deepfake detection
Performance under pressure during the Coronavirus crisis: how a freshly assembled team of AI experts addressed a difficult challenge in unusual circumstances.
In the past 2-years, deepfake videos of Nancy Pelosi, the US Speaker of the House, Facebook’s CEO Mark Zuckerberg and “Game of Thrones” actor Jon Snow have all gone viral and forced many to question deeply what they are seeing and hearing on their screens.
Deepfake, a portmanteau word of “deep learning” and “fake”, refers to AI-generated images made of a digital composite face superimposed on to an existing video (and sometimes audio) of a person. Deepfakes were first mentioned on a Reddit thread in 2017 and have quickly become mainstream in media, politics and popular culture.
Created using open-source machine learning tools, Deepfakes can produce high quality realistic videos. They have been used to create fun memes, but, more sinisterly, they have been used to manipulate, harm individuals, and spread misinformation. They are a major issue for tech and media platform companies that are now facing increasing pressure to remove misleading content.
Indust.ai and the 2020 Deepfake Detection Challenge
In this context, Facebook, Microsoft, AWS and the Partnership on AI joined together in late 2019 to create the Deepfake Detection Challenge on the Kaggle platform, with a sizeable USD 1 Million prize pool, to attract top AI talent from around the globe in a contest to innovate and provide solutions leading to more reliable automated detection of deepfakes.
At about the same time, we launched indust.ai, our private Artificial Intelligence R&D Lab. Our mission is to help European companies transform their core business activities using AI science, expertise and talent.
As a newly constituted company, with none of us having worked on deepfake detection before, this challenge seemed like a good opportunity to test our ability to come together as a “brain pool” in solving a difficult technical issue, and ultimately help us grow as a team. Deepfake detection is a real-world problem and solving it serves our mission to apply AI to make a difference. We also saw our participation as a good opportunity to up-skill in computer vision.
Our final ranking in the global Challenge leaderboard would validate our efforts.
When the challenge closed on 31st March, 2020 indust.ai ranked in the top 10%.
Diving into deepfakes
Initially, it was just one of us analyzing the dataset, exploring state-of-the-art deepfake detection techniques and sharing his insights with the rest of the team. In early February, we ramped up our efforts, and by the end of the month the team was augmented with a part-time mix of AI and computer vision experts, and engineers with less experience in computer vision to make this challenge a learning opportunity. Scaling the team up from one to five people required putting in place specific tools, methods, best practices, and policies to ensure collaboration efficiency towards that short term goal. All the expected collaborative tools were quickly put in place.
Our team organically came together around the challenge, with each decision based on the trust and experience within the group but also supported through exchanges in which more experienced researchers and engineers introduced the team to new concepts, techniques or tools. This method was also reinforced by our culture, which encourages self-management, distributed leadership, empowerment, wholeness and purpose, as described in Frederic Laloux’s “Reinventing organizations” and Simon Sinek’s “Start with Why?”, two inspirations for the indust.ai founding team.
Our expanded efforts quickly paid off, as our ranking in the challenge leaderboard improved consistently.
Then the Coronavirus bomb dropped
With some of us having already had to implement business continuity planning (during SARS and MERS), as soon as we understood the scale of what was happening, we moved to protect our team early. We set up company-wide remote work starting 12th March, 2020, five days before the official French government confinement was announced on 17th March, 2020. Our move was swift, with screens, laptops, power cords, and reserves of dark chocolate being hastily grabbed from the office overnight. Within 48 hours, we had established a regular routine of morning and afternoon stand-up meetings — essential touch points for the team, where we would discuss everything, from client projects to the deepfake challenge: questions our latest submission had thrown up, or who could tackle an unforeseen issue around Google Cloud Platform usage…
The indust.ai brain pool’s performance
When the challenge closed on 31st March 2020, indust.ai placed in the top 10% of all 2,281 competitors. For a team that had only recently come together, with different levels of experience and backgrounds, and competing against global teams from China, the USA, Israel and Europe among others, both from academia and the industry, we are very pleased with our collective performance.
Key to our success was improving our face detection pipeline, introducing a variety of data augmentation techniques, and testing several convolutional neural network architectures. To learn more about our approach take a look at our technical article.
AI will be key for companies to become future-proof and indust.ai is ready to help
Our final ranking serves as tangible proof that our “brain pool” approach to solving new problems works. Our setup is now road-tested, and will be evolving as we continue to focus on our mission: to future-proof companies by bringing them advanced AI science to support their much-needed business transformation.
The work we’ve done on deepfake detection can be directly applied to address problems including those faced by social media networks in their fight against fake news, e-commerce companies struggling with scams, or media companies that must authenticate their sources.
The tools we’ve developed during the challenge, from efficient data augmentation strategies to our own fast face tracking system, will also support the transformation of industries.
To learn more about Computer Vision, or how Artificial Intelligence and Machine Learning can support your business strategy please do contact us: info@indust.ai
You can read more about our technical approach to the Deepfake Challenge here.
Supporting links / References
