Deploying deep learning (DL) projects are becoming increasingly more pervasive at enterprises and startups alike. At Amazon, Machine Learning University (MLU)-trained engineers are taking DL to every aspect of Amazon's businesses, beyond just Amazon Go, Alexa, and Robotics.

Abstract Today’s automobiles leverage powerful sensors and embedded computers to optimize efficiency, safety, and driver engagement. However the complexity of possible inferences using in-car sensor data is not well understood. While we do not know of attempts by automotive manufacturers or makers of after-market components (like insurance dongles) to violate privacy, a key question we ask is: could they (or their collection and later accidental leaks of data) violate a driver’s privacy? In the present study, we experimentally investigate the potential to identify individuals using sensor data snippets of their natural driving behavior. More specifically we record the in-vehicle sensor data on the controllerarea- network (CAN) of a typical modern vehicle (popular 2009 sedan) as each of 15 participants (a) performed a series of maneuvers in an isolated parking lot, and (b) drove the vehicle in traffic along a defined ~ 50 mile loop through the Seattle metropolitan area. We then split the data into training and testing sets, train an ensemble of classifiers, and evaluate identification accuracy of test data queries by looking at the highest voted candidate when considering all possible one-vs-one comparisons. Our results indicate that, at least among small sets, drivers are indeed distinguishable using only incar sensors. In particular, we find that it is possible to differentiate our 15 drivers with 100% accuracy when training with all of the available sensors using 90% of driving data from each person. Furthermore, it is possible to reach high identification rates using less than 8 minutes of training data. When more training data is available it is possible to reach very high identification using only a single sensor (e.g., the brake pedal). As an extension, we also demonstrate the feasibility of performing driver identification across multiple days of data collection

We introduce SensorSift, a new theoretical scheme for balancing utility and privacy in smart sensor applications. At the heart of our contribution is an algorithm which transforms raw sensor data into a 'sifted' representation which minimizes exposure of user defined private attributes while maximally exposing application-requested public attributes. We envision multiple applications using the same platform, and requesting access to public attributes explicitly not known at the time of the platform creation. Support for future-defined public attributes, while still preserving the defined privacy of the private attributes, is a central challenge that we tackle.

We conduct an extensive study of information leakage over the powerline infrastructure from eight televisions (TVs) spanning multiple makes, models, and underlying technologies. In addition to being of scientific interest, our findings contribute to the overall debate of whether or not measurements of residential powerlines reveal significant information about the activities within a home. We find that the power supplies of modern TVs produce discernible electromagnetic interference (EMI) signatures that are indicative of the video content being displayed. We measure the stability of these signatures over time and across multiple instances of the same TV model, as well as the robustness of these signatures in the presence of other noisy electronic devices connected to the same powerline.