AI Security Solutions
Revatron has developed the world’s first smart camera that can identify and disregard adversarial patches. “Adversarial Patches” is a general term describing a strategically placed small patch of pixels capable of misleading machine learning systems to report false classifications. These patches can be easily added or attached to original images by various methods to trick recognition systems. A major concern about adversarial patches is that terrorists may use them to manipulate AI systems for terror attacks, such as fooling autonomous driving machine learning systems to think a red light is green or that a stop sign is on the road when there isn’t one there.
Revatron’s revolutionary cameras can defeat these patches using its real-time AI engine that learns the 3D structure of surrounding objects. Because Revatron’s cameras classify objects based on 3D information rather than simple 2D data used by mainstream machine learning systems, Revatron’s smart cameras see adversarial patches as flat objects without meaningful 3D representations. Thus the cameras classify adversarial patches as a post or a picture with a flat surface and can filter these out, fundamentally resolving the potential threats caused by adversarial patches.
This breakthrough camera is based on Revatron’s DOORs (Direct Object-Oriented Reality system) technology, a real-time AI solution focusing on machine learning and 3D reality modeling. The real-time AI function of Revatron’s camera, called the DOORs camera, is run on a 4W multicore processor with an FPGA accelerator. Revatron has demonstrated the prototype design of the DOORs camera to select cloud and automotive companies and has also showed the conceptual design at the NAB and ISC West trade shows earlier this year.
3D Machine Learning for Video Classification
Revatron has developed the world’s first real-time AI solution for learning 3D depth, motion, and activities of objects in video for the application of video classification. This AI solution consists of Socionext A11 processors and FPGA chips as accelerators. It is a modular design in which each module has two A11 processors and one FPGA chip with a power budget of approximately 20 watts. Each module can process up to 10 full HD video streams in real-time.
Revatron’s real-time AI solution uses triangulation to learn depth and motion of objects within the video based on global motions frame by frame. It can send alerts to an operator if a video matches pre-defined criteria. The module can be trained by an operator to increase its accuracy and proficiency. It is the world’s first real-time AI system capable of learning object behaviors via 3D motion and classification of the video based on object behaviors.
Blocking inappropriate content from live streaming video is a daunting task for social media providers and screening video prior to posting adds complexity to the task. Millions of videos are posted online daily. Real-time filtering of inappropriate content is an urgent issue. Currently available machine learning solutions can only handle 2D images without the ability to learn depth or motion of objects which only can be learned from video as opposed to a single image. Learning behavior is a key advancement for effective video classification.
This breakthrough real-time AI solution is based on Revatron’s DOORs (Direct Object-Oriented Reality system) technology. Revatron’s real-time AI solution is complimentary to all existing machine learning systems. Revatron invites all AI developers to work with us to make an ultimate video classifier to meet every particular need.
View Lidar – A LIDAR Alternative
Revatron has developed the world’s first smart camera enabling automobiles to learn distances and motions of objects just by using the camera. Revatron’s smart camera can also distinguish objects in motion from stationary ones by compensating for the movement of the vehicle. This camera embeds a real-time AI engine with a sub-one millisecond processing time. The real-time feature processing and ultra-low latency learning are optimized for automobile applications.
This breakthrough camera is based on Revatron’s DOORs (Direct Object-Oriented Reality system) technology, a real-time AI solution focusing on machine learning and 3D reality modeling. The camera can support from one to three camera inputs for 3D measurements. The DOORs camera’s passive design learns 3D depth information from the triangulation of multiple synchronized camera inputs. When using only a single camera input, the camera uses its own travelling path as the basis for triangulation.
The automobile industry has been looking for a LIDAR replacement due to its limitations in affordable mass production applications. Basically, LIDAR is a spinning radar system that requires a vantage point to survey its surroundings without blind spots. However, LIDAR is large, heavy, expensive, and not suited for high volume production. A typical autonomous driving system using LIDAR may require a 2KW computer to build a limited 3D model around the car in about 1.5 seconds.
Unlike LIDAR, the DOORs camera is a completely passive device that can be as small as a single compact drive recorder and mounted anywhere in a vehicle as an accessory to provide voice warnings or information on the surrounding area. The DOORs camera learns the 3D structures of the surroundings from only the camera inputs that don’t emit any lights or signals. The DOORs camera can provide the precise movements of surrounding objects if both GPS and speed data are available.
Low Latency Video and Data Transmission System
Transmitting video or data over the Internet within 50 Milliseconds
A real-time AI powered end-to-end network transmission system can deliver video or data within 50 milliseconds across the Internet. It is achieved by changing network packet sizes and session settings after learning the Internet traffic few hundred times per second.