Protopia AI offers an exclusive solution for an overlooked challenge, inference privacy and data protection to enable inter- and intra-enterprise data sharing and securing inference services against data leaks.

Data used in inference services contains a staggering amount of privileged and private information across many industries such as finance, healthcare, insurance, voice assistants, smart speakers, surveillance systems, and others. The interwoven mix of data poses significant risks for businesses and their customers. While data is protected at rest and in motion through encryption, it will be exposed during inference as that data needs to be processed in an un-encrypted fashion.

Protopia AI addresses this structural gap in inference privacy using a novel obfuscation technology, which leverages gradient mechanisms to find stochastic data transformations that obfuscate the data while also keeping the inference service highly performant.

This solution for Confidential Inference–demoed here–is part of Protopia AI’s suite of AI data and model transformations. These transformations protect access to the data and integrity of the AI models in an automated fashion. Protopia’s solutions reduce restrictions facing data sharing for AI, enhance data security and privacy for AI and help identify vulnerabilities to adversarial attacks, as well as protect models from inversion attacks.

Protopia AI’s solutions significantly shrink the attack surface at the data level before compute starts. As such, Protopia accelerates the deployment process of AI, minimizes exposure to leakage of sensitive data and models, and prevents unintended inferences.

One of the most challenging aspects of online disinformation is the overwhelming volume of content that is published on social media platforms. For example, hundreds of thousands of images and videos are uploaded to Facebook every minute. Organizing and analyzing this volume of content in the hope of detecting disinformation campaigns in (near) real-time is impossible for humans without the assistance of automated AI tools. This problem is especially pertinent in young and struggling democracies whose traditional media organizations lack the ability to keep pace with the explosion of deep-fake, manipulated, altered or plainly-fake online media. In an effort to provide such capacity, we have developed a real-time social media manipulation detection and analysis system called MEWS (Misinformation Early Warning System).

This system combines work in digital forensics, computer vision, graph analysis, and media studies to accomplish three specific tasks: (1) MEWS ingests enormous amounts of images and video from various social media platforms (e.g., Facebook, Instagram, Twitter, Telegram) using keyword targets provided by partner media organizations from across the world; (2) MEWS employs state-of-the-art AI systems to detect and extract faces, objects, text (including meme-text), image features, and any potential manipulations from the visual content; and (3) MEWS constructs a media-graph which pairs similar sub-images, objects, and manipulations for display in an interactive, easily-navigable, and searchable user interface.

We offer a demonstration of MEWS' organizational and analytic capabilities using tens of millions of images (and other media) collected from several social media platforms (Facebook, Instagram, and Twitter) in the Indonesian context.

Word vector embeddings have been shown to contain and amplify biases in data they are extracted from. Consequently, many techniques have been proposed to identify, mitigate, and attenuate these biases in word representations. In this tutorial, we will review a collection of state-of-the-art debiasing techniques. To aid this, we provide an open source web-based visualization tool and offer hands-on experience in exploring the effects of these debiasing techniques on the geometry of high-dimensional word vectors. To help understand how various debiasing techniques change the underlying geometry, we decompose each technique into interpretable sequences of primitive operations, and study their effect on the word vectors using dimensionality reduction and interactive visual exploration.

Developing efficient data-driven applications, especially using deep learning, requires access to large and diverse datasets. However, sharing and collecting sensitive data is extremely challenging due to privacy, ethical, and legal concerns. To address these challenges, we present TripleBlind, a practical privacy-preserving framework for creating and consuming data-driven applications from decentralized data and algorithms. TripleBlind provides a set of automated, high-level APIs that enable (1) extracting conclusions from remote data without moving it outside the owner's firewall, (2) training sophisticated AI models from decentralized data, and (3) consuming trained models for secure and efficient inference-as-a-service without compromising the privacy of either the model or the data. We focus in this tool demo on two tasks: First, we train a ResNet34 model using decentralized medical image data over the public Internet without "seeing" the raw data. Second, we utilize our secure multi-party computation protocol to run real-time inference using the trained model over the public Internet.

Millions of people around the world can not access content on the Web because most of the content is not readily available in their language. Machine translation (MT) systems have the potential to change this for many languages. Current MT systems provide very accurate results for high resource language pairs, e.g., German and English. However, for many low resource languages, MT is still under active research. The key challenge is lack of datasets to build these systems.

We present Lesan, an MT system for low resource languages. Our pipeline solves the key bottleneck to low resource MT by leveraging online and offline sources, a custom OCR system for Ethiopic and an automatic alignment module. The final step in the pipeline is a sequence to sequence model that takes parallel corpus as input and gives us a translation model. Lesan's translation model is based on the Transformer architecture. After constructing a base model, back translation, is used to leverage monolingual corpora.

Currently Lesan supports translation to and from Tigrinya, Amharic and English. We perform extensive human evaluation and show that Lesan outperforms state-of-the-art systems such as Google Translate and Microsoft Translator across all six pairs. Lesan is freely available and has served more than 10 million translations so far. At the moment, there are only 213 Tigrinya and 14,964 Amharic Wikipedia articles. We believe that Lesan will contribute towards democratizing access to the Web through MT for millions of people.