Delighted to open today’s Washington Post newsletter to see a prominent headline about audio deepfakes- with quotes from Hany and myself in it!
Check out the article here: https://www.washingtonpost.com/technology/interactive/2024/ai-voice-detection-trump-harris-deepfake-election/
Our piece about the AI-deepfake of Bad Bunny is out in the Berkeley Science Review now.
On 17th July 2024, I was delighted to be able to speak as a keynote technical presenter at the Alan Turing Institute Women in AI Security Workshop.
I was invited to share my thoughts with Vox about recent developments in the entertainment industry revolving around FKA twigs, Scarlett Johansson and the looming dawn of audio deepfakes.
Was great to be a part of this recent NPR piece on voice cloning. I got to talk to Huo Jingnan about all things VC detection, and how there’s no ‘silver bullet’ from machine learning to fix this problem.
My PI, Hany Farid, and I were delighted to join the discussion about AI voice cloning technologies at The White House in January.
How do you detect a cloned voice? The simple answer… deep learning. Hugely enjoyed presenting our different detection algorithms and the relative benefits of each at the 2023 IEEE International Workshop on Information and Forensics (WIFS) in Nuremburg, Germany.
Had a fantastic day leading the AI & cybersecurity session for the U.S. Department of State #TechWomen23 initiative. I presented an overview of the AI landscape at present; ranging from deepfakes and misinformation to the potential threat of cyber-fuelled nuclear warfare.
Selected student facilitator for the Berkeley Risk and Security Lab & OpenAI workshop on Confidence-Building Measures for Artificial Intelligence (Jan 2023, paper released August 2023).
Recently, I was fortunate to work with Romit Barua and Gautham Koorma on a project exploring the relative performances of computation approaches to cloned-voice detection. Synthetic-voice cloning technologies have seen significant advances in recent years, giving rise to a range of potential harms.
Excited, shocked (!) and honored to share that our work in deepfake detection was recognized at The Nobel Prize summit by the Digital Public Goods Alliance and United Nations UNDP as part of their campaign in combatting disinformation.
We were invited to attend the summit and participate in the most enriching and enlightening conversations about information integrity. We met world leaders, Laureates, policymakers, and technologists from the international community who are creating real change in their fields. We were also invited to the Royal Norwegian Embassy to present our work.
We were selected as one of an array of open-source innovations aiming to enhance information integrity worldwide. More info at bit.ly/Info-Pollution #nobelprizesummit #nobelprize #unitednations. Thank you to our lab Romit Barua Gautham Koorma Hany Farid UC Berkeley School of Information, and a special shout out to Nick Merrill for sharing this amazing opportunity.
Hany & I have been working for over a year devising a study to examine how well we can estimate height and weight from a single image. We compare state-of-the-art AI, computer vision, and 3D modeling techniques with estimations from experts and non-experts. The results are surprising.
Classical computer science techniques for the re-projection of an image onto a non-planar surface have been extended to modeling the geometry of the human eye, as seen in Ko and Nayar (2004). Yet, there are few examples of the application of these methods to the analysis of corneal reflections in deep fake detection. The work of Hu et al. (2021) explored inconsistencies in inter-eye reflections, but focussed solely on specular highlights as opposed to the reflected scene itself. This work aims to build upon these two approaches in order to investigate whether there are discernible differences between the corneal reflections of real vs. deep-fake human portraits.
The findings from this approach provide several key contributions:
A framework for automatically extracting the corneal surface of a human in a single image and generating iris-based features.
For two publicly-available datasets, we ultimately find that there are no significant discernible differences between real and fake corneal reflections (based on the performance of three classification models).
We hypothesize that this lack of differentiation between real and fake portraits is primarily due to insufficient image quality once cropped (which typically reduces a 1024x1024px photo down to approximately 50x50px). This clearly impacts the pixel representation of the corneal surface and introduces noise that distorts the extracted features. The noisy nature of the dataset is evident from the outputs of Principal Component Analysis, which indicates that 10 components are required to capture just 75% of the variation in the data.
In spite of this, the feature extraction process highlights several features of interest that, qualitatively, can be seen to differentiate between real and fake portraits. This includes the shape of the pupil, which, for the majority of humans, should form a near-perfect circle shape when facing toward the camera. Yet, in several GAN-synthesised images from our dataset, we observe inconsistencies such as straight edges, corners, and even no clear pupil at all.
Additionally, edges observed in the reflections of real corneal surfaces (particularly in outdoor or natural light settings) exhibit discernable properties such as objects or information about the scene. By contrast, several deep-fake corneal reflections were observed to contain noise and abstract lighting, not resembling a coherent scene or object. However, we hypotheise that while evident to a human, the edge count feature was skewed by noise in the low resolution cropped images, thus not providing a strong predictor.
Analysis: https://github.com/sbarrington/corneal-reflections/tree/main/01%20Paper
Public Dataset: https://github.com/sbarrington/cornea-iris-dataset (DOI: 10.5281/zenodo.7396604)
Analysis snapshots:
Non-Fungible Tokens (NFTs), digital certificates of ownership for virtual art, have until recently been traded on a highly lucrative and speculative market. Yet, an emergence of misconceptions, along with a sustained market downtime, are calling the value of NFTs into question.
Since late last year, I have been collaborating on two very exciting projects regarding the feasibility of applications for Web3 technologies as part of the University of California Big Ideas contest.
In April, I attended the Stanford University CodeX Blockchain Group's DAOs and Systems for Resilient Societies conference as a collaborator with my colleagues from the Open Earth Foundation (OEF). OEF is a non-profit research organisation whom I have worked with on several research projects, examining the feasibility and security implications of decentralised technologies for climate applications.
Defined as a conceptual model to enable the consideration of ‘how power, oppression, resistance, privilege, penalties, benefits and harms are systematically distributed’, the matrix of domination is a term coined by feminist scholar Patricia Hill Collins in her book Black Feminist Thought. Sociotechnical, data-driven systems can, in particular, be considered to enforce & proliferate this structure; which constitutes the thesis explored by Costanza-Chocks’s article written in 2017. Arguably, the most notable example of such interlocking systems of oppression can be observed in social media moderation algorithms, which are typically combined human-algorithmic approaches to curating and silencing online user-generated content.
Over the past 2 decades, Digital Platforms (DPs), including Facebook, Google, Amazon and Apple, have risen from insignificant start- ups to a dominating set of firms with a combined 4 trillion dollar market capitalisation1. These now ubiquitous DPs have revolutionised many aspects of everyday life, from working and studying to communicating and dating. In recognition of this unprecedented rise of commercial power, legal and academic scholarship has begun to revisit the concepts of monopolistic market behaviours and the subsequent potential for both economic and political influence.
Cryptocurrency and Blockchain technologies are fast becoming areas of public interest across a breadth of diverse domains, with novel applications ranging from financial services to state politics. In particular, promising future use cases of these decentralised technologies aim to re-define the concept of value, and potentially enable a new wave of accessible investment opportunities for the general public. One such example of this is through Non-Fungible Tokens (NFTs), which act as digital certificates of ownership for a given piece of digital content. These stamps are stored as tokens on a public Blockchain (such as Ethereum), which uses complex hashing algorithms to ensure that each token is unique, tamper-proof and publicly-visible. This adds value to a range of digital content & artworks through enabling a single source of truth relating to its ownership and historic activity.
As a component of the INFO-272 Research course at the School of Information, I lead an independent qualitative research study in order to explore the perception of threat in a range of online abuse situations through the contexts, themes and experience of five self-identifying women. Online abuse is becoming an increasingly prevalent issue in modern day society, with 41% of Americans having experienced online harassment in some capacity in 2021. People who identify as women, in particular, can be subjected to a wide range of abusive behaviour online, with gender-specific experiences cited broadly in recent literature across fields such as blogging, politics and journalism.
