In this paper we consider the problem of "end-to-end" digital camera identification by considering sequence of images obtained from the cameras. The problem of digital camera identification is harder than the problem of identifying its analog counterpart since the process of analog to digital conversion smooths out the intrinsic noise in the analog signal. However it is known that identifying a digital camera is possible by analyzing the camera's intrinsic sensor artifacts that are introduced into the images/videos during the process of photo/video capture. It is known that such methods are computationally intensive requiring expensive pre-processing steps. In this paper we propose an end-to-end deep feature learning framework for identifying cameras using images obtained from them. We conduct experiments using three custom datasets: the first containing two cameras in an indoor environment where each camera may observe different scenes having no overlapping features, the second containing images from four cameras in an outdoor setting but where each camera observes scenes having overlapping features and the third containing images from two cameras observing the same checkerboard pattern in an indoor setting. Our results show that it is possible to capture the intrinsic hardware signature of the cameras using deep feature representations in an end-to-end framework. These deep feature maps can in turn be used to disambiguate the cameras from each another. Our system is end-to-end, requires no complicated pre-processing steps and the trained model is computationally efficient during testing, paving a way to have near instantaneous decisions for the problem of digital camera identification in production environments. Finally we present comparisons against the current state-of-the-art in digital camera identification which clearly establishes the superiority of the end-to-end solution.

Deep Neural Networks (DNNs) have become the tool of choice for machine learning practitioners today. One important aspect of designing a neural network is the choice of the activation function to be used at the neurons of the different layers. In this work, we introduce a four-output activation function called the Reflected Rectified Linear Unit (RReLU) activation which considers both a feature and its negation during computation. Our activation function is "sparse", in that only two of the four possible outputs are active at a given time. We test our activation function on the standard MNIST and CIFAR-10 datasets, which are classification problems, as well as on a novel Computational Fluid Dynamics (CFD) dataset which is posed as a regression problem. On the baseline network for the MNIST dataset, having two hidden layers, our activation function improves the validation accuracy from 0.09 to 0.97 compared to the well-known ReLU activation. For the CIFAR-10 dataset, we use a deep baseline network that achieves 0.78 validation accuracy with 20 epochs but overfits the data. Using the RReLU activation, we can achieve the same accuracy without overfitting the data. For the CFD dataset, we show that the RReLU activation can reduce the number of epochs from 100 (using ReLU) to 10 while obtaining the same levels of performance.

In this paper we present a large scale, passive positioning system that can be used for approximate localization in Global Positioning System (GPS) denied/spoofed environments. This system can be used for detecting GPS spoofing as well as for initial position estimation for input to other GPS free positioning and navigation systems like Terrain Contour Matching (TERCOM). Our Location inference through Frequency Modulation (FM) Signal Integration and estimation (LoSI) system is based on broadcast FM radio signals and uses Received Signal Strength Indicator (RSSI) obtained using a Software Defined Radio (SDR). The RSSI thus obtained is used for indexing into an estimated model of expected FM spectrum for the entire United States. We show that with the hardware for data acquisition, a single point resolution of around 3 miles and associated algorithms, we are capable of positioning with errors as low as a single pixel (more precisely around 0.12 mile). The algorithm uses a large-scale model estimation phase that computes the expected FM spectrum in small rectangular cells (realized using geohashes) across the Contiguous United States (CONUS). We define and use Dominant Channel Descriptor (DCD) features, which can be used for positioning using time varying models. Finally we use an algorithm based on Euclidean nearest neighbors in the DCD feature space for position estimation. The system first runs a DCD feature detector on the observed spectrum and then solves a subset query formulation to find Inference Candidates (IC). Finally, it uses a simple Euclidean nearest neighbor search on the ICs to localize the observation. We report results on 1500 points across Florida using data and model estimates from 2015 and 2017. We also provide a Bayesian decision theoretic justification for the nearest neighbor search.