We introduce a new computer aided detection and diagnosis system for lung cancer screening with low-dose CT scans that produces meaningful probability assessments. Our system is based entirely on 3D convolutional neural networks and achieves state-of-the-art performance for both lung nodule detection and malignancy classification tasks on the publicly available LUNA16 and Kaggle Data Science Bowl challenges. While nodule detection systems are typically designed and optimized on their own, we find that it is important to consider the coupling between detection and diagnosis components. Exploiting this coupling allows us to develop an end-to-end system that has higher and more robust performance and eliminates the need for a nodule detection false positive reduction stage. Furthermore, we characterize model uncertainty in our deep learning systems, a first for lung CT analysis, and show that we can use this to provide well-calibrated classification probabilities for both nodule detection and patient malignancy diagnosis. These calibrated probabilities informed by model uncertainty can be used for subsequent risk-based decision making towards diagnostic interventions or disease treatments, as we demonstrate using a probability-based patient referral strategy to further improve our results.

Increasing numbers of software vulnerabilities are discovered every year whether they are reported publicly or discovered internally in proprietary code. These vulnerabilities can pose serious risk of exploit and result in system compromise, information leaks, or denial of service. We leveraged the wealth of C and C++ open-source code available to develop a largescale function-level vulnerability detection system using machine learning. To supplement existing labeled vulnerability datasets, we compiled a vast dataset of millions of open-source functions and labeled it with carefully-selected findings from three different static analyzers that indicate potential exploits. Using these datasets, we developed a fast and scalable vulnerability detection tool based on deep feature representation learning that directly interprets lexed source code. We evaluated our tool on code from both real software packages and the NIST SATE IV benchmark dataset. Our results demonstrate that deep feature representation learning on source code is a promising approach for automated software vulnerability detection.

Motivated by the problem of automated repair of software vulnerabilities, we propose an adversarial learning approach that maps from one discrete source domain to another target domain without requiring paired labeled examples or source and target domains to be bijections. We demonstrate that the proposed adversarial learning approach is an effective technique for repairing software vulnerabilities, performing close to seq2seq approaches that require labeled pairs. The proposed Generative Adversarial Network approach is application-agnostic in that it can be applied to other problems similar to code repair, such as grammar correction or sentiment translation.

We consider the problem of fusing probability scores from a set of classifiers to estimate a final fused probability score. Our interest is in scenarios where the classifiers are statistically dependent. To that end, we propose a new classifier fusion approach that is data driven and founded on the statistical theory of copulas. Numerical results with both simulated and real data show that our copula based classifier fusion approach produces better probability scores than individual classifiers and outperforms existing probability score fusion approaches.

Automatic modulation classification (AMC) with multiple sensors is a challenging problem when the channel conditions are unknown at the receiver. In this paper, using the Markov chain Monte Carlo (MCMC) approach, we develop a novel algorithm for AMC when the amplitude and phase of the channel gains are unknown. Using sampling techniques, we marginalize over the unknown channel parameters that follow a certain probability distribution. This improves the estimate of the a posteriori distribution of the modulation formats, thereby improving the overall classification performance. Further, to overcome the problem of local extrema traps encountered in sampling algorithms, we introduce the idea of adding artificial noise beyond a certain threshold of signal-to-noise (SNR). This improves the performance of the sampling based AMC algorithm in the high SNR regime. Simulation results and comparisons are provided to show the efficiency of the proposed algorithm over the most related works in the literature.

We employ Permutation Trellis Code (PTC) based multi-level Frequency Shift Keying signaling to mitigate the impact of Primary Users (PUs) on the performance of Secondary Users (SUs) in Cognitive Radio Networks (CRNs). The PUs are assumed to be dynamic in that they appear intermittently and stay active for an unknown duration. Our approach is based on the use of PTC combined with multi-level FSK modulation so that an SU can improve its data rate by increasing its transmission bandwidth while operating at low power and not creating destructive interference for PUs. We evaluate system performance by obtaining an approximation for the actual Bit Error Rate (BER) using properties of the Viterbi decoder and carry out a thorough performance analysis in terms of BER and throughput. The results show that the proposed coded system achieves i) robustness by ensuring that SUs have stable throughput in the presence of heavy PU interference and ii) improved resiliency of SU links to interference in the presence of multiple dynamic PUs.

In this paper, we address the constrained resource allocation problems arising in the context of spectrum sharing in cognitive radio networks utilizing a multi-dimensional formulation. Given the activity of the primary users (PUs), we consider multiple objectives and constraints, viz., sum rate, fairness, number of active secondary users (SUs), power consumption, and quality of service requirements (of both PUs and SUs). The three dimensions for the optimization task are the assignment of power, frequency, and antenna directionality to various SUs. Efficient heuristic algorithms are developed for five variations of the NP-hard optimization problems. Solution quality tradeoffs are shown for three algorithms, viz., convex relaxation with tree pruning, convex relaxation with gradual removal, and a genetic algorithm (GA); results show that the GA provides a reasonable balance between solution quality and computational effort. The multi-objective problems are solved using a modification of the NSGA-II evolutionary algorithm, obtaining a set of Pareto-optimal solutions under computational constraints.

In this paper, we consider distributed maximum likelihood (ML) classification of digital amplitude-phase modulated signals using multiple sensors that observe the same sequence of unknown symbol transmissions over nonidentical flat blockfading Gaussian noise channels. A variant of the expectation-maximization (EM) algorithm is employed to obtain the ML estimates of the unknown channel parameters and compute the global log-likelihood of the observations received by all the sensors in a distributed manner by means of an average consensus filter. This procedure is repeated for all candidate modulation formats in the reference library, and a classification decision, which is available at any of the sensors in the network, is declared in favor of the modulation with the highest log-likelihood score. The proposed scheme improves the classification accuracy by exploiting the signal-to-noise ratio (SNR) diversity in the network while restricting the communication to a small neighborhood of each sensor. Numerical examples show that the proposed distributed EM-based classifier can achieve the same classification performance as that of a centralized classifier, which has all the sensor measurements, for a wide range of SNR values.

In this paper, we consider the problem of automatic modulation classification with multiple sensors in the presence of unknown time offset, phase offset and received signal amplitude. We develop a novel hybrid maximum likelihood (HML) classification scheme based on a generalized expectation maximization (GEM) algorithm. GEM is capable of finding ML estimates numerically that are extremely hard to obtain otherwise. Assuming a good initialization technique is available for GEM, we show that the classification performance (in terms of the probability of error) can be greatly improved with multiple sensors compared to that with a single sensor, especially when the signal-to-noise ratio (SNR) is low. We further demonstrate the superior performance of our approach when simulated annealing (SA) with uniform as well as nonuniform grids is employed for initialization of GEM in low SNR regions. The proposed GEM based approach employs only a small number of samples (in the order of hundreds) at a given sensor node to perform both time and phase synchronization, signal power estimation, followed by modulation classification. We provide simulation results to show the efficiency and effectiveness of the proposed algorithm.

We address the problem of discovering unknown digital amplitude-phase modulations over block-fading additive noise channels. The proposed method uses the iterative Richardson-Lucy algorithm to determine the distribution of the transmitted symbols, which completely characterizes the underlying signal constellation. The decoding of the received signals can then be carried out based on the estimate of the signal constellation. An important application of the proposed method is to construct a modulation dictionary in an offline manner prior to performing any type of real time classification, thereby improving the performance of the automatic modulation classification algorithms proposed in the literature.

In this paper, we consider the problem of linear modulation classification in the presence of unknown time offset, phase offset and received signal amplitude. We develop a novel hybrid maximum likelihood (HML) approach based on a Generalized Expectation Maximization (GEM) algorithm [1]. Our approach is applicable to all QAM and PSK modulations, and it does not require any assumptions on the received signal-to-noise ratio (SNR). The GEM algorithm provides a tractable procedure to obtain maximum likelihood (ML) estimates which are extremely hard to obtain otherwise. Moreover, our approach employs only a small number of samples (in the order of hundreds) to perform both time and phase synchronization, signal power estimation, followed by modulation classification. The proposed approach also enables maximum a posteriori (MAP) decoding of the unknown constellation symbol sequence as a by-product of the GEM algorithm. We provide simulation results that show that the proposed approach provides excellent classification performance.

We analyze the potential for using multiple transmission dimensions in a cognitive radio network (CRN) in terms of the probability of successful communications. We consider a random CRN where the users are randomly distributed in an area of interest. The users can be mobile. We derive the successful communication probability (SCP) with respect to transmit power for different density of primary users and secondary users by including different transmission dimensionalities such as time, frequency and antenna directionality. It is shown that using multiple transmission dimensions improves the SCP significantly. The advantage of using directional antennas for spatial reuse and increased range is also shown in terms of improved SCP.

We propose a modified Bayesian Cramér-Rao lower bound (BCRLB) for nonlinear tracking applications where the prediction distribution conditioned on past measurements is used as the prior. The novelty of the proposed modified BCRLB comes from the fact that it utilizes past measurements, therefore it is specific to the current realization of the track which makes it a useful online tool that can be used for real-time sensor management. The computation of our proposed modified BCRLB is not analytically tractable except under very restricted conditions. Therefore, we also develop a particle based numerical computation method for our modified BCRLB so that this new bound can be easily calculated in real-time using the particles already available from the underlying particle filter which is used to track the target. We show by simulations that our developed numerical computation method approaches to its true analytical value as the number of particles in the particle filter increases.

The problem of dynamic bit allocation for target tracking is investigated in this paper under a total sum rate constraint in sensor networks. Bits are dynamically allocated to sensors in such a way that a cost function, which is based on the Cramér-Rao lower bound evaluated at the predicted target state, is minimized. The optimal solution to this problem, namely joint bit allocation and local quantizer design, is computationally prohibitive and not realistic for real-time online implementation. Instead, a two-step optimization procedure is proposed. First, the best time independent quantizers are obtained offline by maximizing the average Fisher information about the signal amplitude, for different number of bits. With the time independent quantizers, the generalized Breiman, Friedman, Olshen, and Stone (BFOS) algorithm is employed to dynamically assign bits to sensors. Simulation results show that with the same or even less sum bit rate, the proposed dynamic bit allocation approach leads to significantly improved tracking performance, compared with the static bit allocation approach where each sensor is allocated with equal number of bits.

We study the parametric distributed estimation problem using a wireless sensor network (WSN) where each sensor observes an unknown scalar parameter, quantizes its observation and sends its quantized observation to a fusion center via fading and noisy communication channels. We propose to incorporate channel statistics rather than the instantaneous channel state information (CSI) into the maximum likelihood (ML) formulation and show that the resulting likelihood function is strictly log-concave almost surely with a change of variable provided that at least one of the communication channels between the sensors and the fusion center has nonzero capacity. We also investigate the effects of channel layer on the sensor threshold design and show that the threshold design problem is coupled with the channel layer and the sensor signal-to-noise ratio (SNR) only for nonsymmetric channels. Our formulation is very general in the sense that no assumptions are made about the physical layer in terms of the modulation schemes and the reception techniques.

In this paper, we design a novel energy-detection based receiver architecture to detect unknown UWB signals in a strong narrowband interference (NBI) environment. Designed receiver is capable of suppressing NBI at low cost without any need for searching its frequency location. This is made possible by preprocessing the received signal using a cascaded nonlinear energy operator followed by a high-pass filter before regular energy detection.

In this paper, we propose a new maximumlikelihood (ML) target location estimator which uses quantized sensor data and wireless channel statistics in a wireless sensor network. The novelty of our approach comes from the fact that imperfect channel statistics between wireless sensors and the fusion center are incorporated in the localization algorithm. We call this approach “channel-aware target localization”. Furthermore, we derive the Cramer-Rao lower bound as a performance bound for our channel-aware ML estimator. Simulation results are presented to show that the performance of the channel-aware ML location estimator is quite close to its theoretical performance bound even with relatively small number of sensors and it has superior performance compared to that of the channel-unaware ML estimator.