Publications
A spectral framework for anomalous subgraph detection
August 15, 2015
Journal Article
Published in:
IEEE Trans. Signal Process., Vol. 63, No. 16, 15 August 2015, 4191-4206.
Topic:
R&D area:
Summary
A wide variety of application domains is concerned with data consisting of entities and their relationships or connections, formally represented as graphs. Within these diverse application areas, a common problem of interest is the detection of a subset of entities whose connectivity is anomalous with respect to the rest of the data. While the detection of such anomalous subgraphs has received a substantial amount of attention, no application-agnostic framework exists for analysis of signal detectability in graph-based data. In this paper, we describe a framework that enables such analysis using the principal eigenspace of a graph's residuals matrix, commonly called the modularity matrix in community detection. Leveraging this analytical tool, we show that the framework has a natural power metric in the spectral norm of the anomalous subgraph's adjacency matrix (signal power) and of the background graph's residuals matrix (noise power). We propose several algorithms based on spectral properties of the residuals matrix, with more computationally expensive techniques providing greater detection power. Detection and identification performance are presented for a number of signal and noise models, including clusters and bipartite foregrounds embedded into simple random backgrounds, as well as graphs with community structure and realistic degree distributions. The trends observed verify intuition gleaned from other signal processing areas, such as greater detection power when the signal is embedded within a less active portion of the background. We demonstrate the utility of the proposed techniques in detecting small, highly anomalous subgraphs in real graphs derived from Internet traffic and product co-purchases.
Summary
A wide variety of application domains is concerned with data consisting of entities and their relationships or connections, formally represented as graphs. Within these diverse application areas, a common problem of interest is the detection of a subset of entities whose connectivity is anomalous with respect to the rest of...
READ MORE
Iris biometric security challenges and possible solutions: for your eyes only? Using the iris as a key
August 13, 2015
Journal Article
Published in:
IEEE Signal Process. Mag., Vol. 32, No. 5, September 2015, pp. 42-53.
Summary
Biometrics were originally developed for identification, such as for criminal investigations. More recently, biometrics have been also utilized for authentication. Most biometric authentication systems today match a user's biometric reading against a stored reference template generated during enrollment. If the reading and the template are sufficiently close, the authentication is considered successful and the user is authorized to access protected resources. This binary matching approach has major inherent vulnerabilities. An alternative approach to biometric authentication proposes to use fuzzy extractors (also known as biometric cryptosystems), which derive cryptographic keys from noisy sources, such as biometrics. In theory, this approach is much more robust and can enable cryptographic authorization. Unfortunately, for many biometrics that provide high-quality identification, fuzzy extractors provide no security guarantees. This gap arises in part because of an objective mismatch. The quality of a biometric identification is typically measured using false match rate (FMR) versus false nonmatch rate (FNMR). As a result, biometrics have been extensively optimized for this metric. However, this metric says little about the suitability of a biometric for key derivation. In this article, we illustrate a metric that can be used to optimize biometrics for authentication. Using iris biometrics as an example, we explore possible directions for improving processing and representation according to this metric. Finally, we discuss why strong biometric authentication remains a challenging problem and propose some possible future directions for addressing these challenges.
Summary
Biometrics were originally developed for identification, such as for criminal investigations. More recently, biometrics have been also utilized for authentication. Most biometric authentication systems today match a user's biometric reading against a stored reference template generated during enrollment. If the reading and the template are sufficiently close, the authentication is...
READ MORE
Temporal and multi-source fusion for detection of innovation in collaboration networks
July 6, 2015
Conference Paper
Published in:
Proc. of the 18th Int. Conf. On Information Fusion, 6-9 July 2015.
Topic:
R&D area:
Summary
A common problem in network analysis is detecting small subgraphs of interest within a large background graph. This includes multi-source fusion scenarios where data from several modalities must be integrated to form the network. This paper presents an application of novel techniques leveraging the signal processing for graphs algorithmic framework, to well-studied collaboration networks in the field of evolutionary biology. Our multi-disciplinary approach allows us to leverage case studies of transformative periods in this scientific field as truth. We build on previous work by optimizing the temporal integration filters with respect to truth data using a tensor decomposition method that maximizes the spectral norm of the integrated subgraph's adjacency matrix. We also demonstrate that we can mitigate data corruption via fusion of different data sources, demonstrating the power of this analysis framework for incomplete and corrupted data.
Summary
A common problem in network analysis is detecting small subgraphs of interest within a large background graph. This includes multi-source fusion scenarios where data from several modalities must be integrated to form the network. This paper presents an application of novel techniques leveraging the signal processing for graphs algorithmic framework...
READ MORE
Deep neural network approaches to speaker and language recognition
April 21, 2015
Journal Article
Published in:
IEEE Signal Process. Lett., Vol. 22, No. 10, October 2015, pp. 1671-5.
Topic:
R&D area:
Summary
The impressive gains in performance obtained using deep neural networks (DNNs) for automatic speech recognition (ASR) have motivated the application of DNNs to other speech technologies such as speaker recognition (SR) and language recognition (LR). Prior work has shown performance gains for separate SR and LR tasks using DNNs for direct classification or for feature extraction. In this work we present the application for single DNN for both SR and LR using the 2013 Domain Adaptation Challenge speaker recognition (DAC13) and the NIST 2011 language recognition evaluation (LRE11) benchmarks. Using a single DNN trained for ASR on Switchboard data we demonstrate large gains on performance in both benchmarks: a 55% reduction in EER for the DAC13 out-of-domain condition and a 48% reduction in Cavg on the LRE11 30 s test condition. It is also shown that further gains are possible using score or feature fusion leading to the possibility of a single i-vector extractor producing state-of-the-art SR and LR performance.
Summary
The impressive gains in performance obtained using deep neural networks (DNNs) for automatic speech recognition (ASR) have motivated the application of DNNs to other speech technologies such as speaker recognition (SR) and language recognition (LR). Prior work has shown performance gains for separate SR and LR tasks using DNNs for...
READ MORE
Planted clique detection below the noise floor using low-rank sparse PCA
April 19, 2015
Conference Paper
Published in:
Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing, ICASSP, 19-24 April 2015.
Topic:
R&D area:
Summary
Detection of clusters and communities in graphs is useful in a wide range of applications. In this paper we investigate the problem of detecting a clique embedded in a random graph. Recent results have demonstrated a sharp detectability threshold for a simple algorithm based on principal component analysis (PCA). Sparse PCA of the graph's modularity matrix can successfully discover clique locations where PCA-based detection methods fail. In this paper, we demonstrate that applying sparse PCA to low-rank approximations of the modularity matrix is a viable solution to the planted clique problem that enables detection of small planted cliques in graphs where running the standard semidefinite program for sparse PCA is not possible.
Summary
Detection of clusters and communities in graphs is useful in a wide range of applications. In this paper we investigate the problem of detecting a clique embedded in a random graph. Recent results have demonstrated a sharp detectability threshold for a simple algorithm based on principal component analysis (PCA). Sparse...
READ MORE
Robust face recognition-based search and retrieval across image stills and video
April 14, 2015
Conference Paper
Published in:
HST 2015, IEEE Int. Symp. on Technologies for Homeland Security, 14-16 April 2015.
Topic:
R&D area:
Summary
Significant progress has been made in addressing face recognition channel, sensor, and session effects in both still images and video. These effects include the classic PIE (pose, illumination, expression) variation, as well as variations in other characteristics such as age and facial hair. While much progress has been made, there has been little formal work in characterizing and compensating for the intrinsic differences between faces in still images and video frames. These differences include that faces in still images tend to have neutral expressions and frontal poses, while faces in videos tend to have more natural expressions and poses. Typically faces in videos are also blurrier, have lower resolution, and are framed differently than faces in still images. Addressing these issues is important when comparing face images between still images and video frames. Also, face recognition systems for video applications often rely on legacy face corpora of still images and associated meta data (e.g. identifying information, landmarks) for development, which are not formally compensated for when applied to the video domain. In this paper we will evaluate the impact of channel effects on face recognition across still images and video frames for the search and retrieval task. We will also introduce a novel face recognition approach for addressing the performance gap across these two respective channels. The datasets and evaluation protocols from the Labeled Faces in the Wild (LFW) still image and YouTube Faces (YTF) video corpora will be used for the comparative characterization and evaluation. Since the identities of subjects in the YTF corpora are a subset of those in the LFW corpora, this enables an apples-to-apples comparison of in-corpus and cross-corpora face comparisons.
Summary
Significant progress has been made in addressing face recognition channel, sensor, and session effects in both still images and video. These effects include the classic PIE (pose, illumination, expression) variation, as well as variations in other characteristics such as age and facial hair. While much progress has been made, there...
READ MORE
Global pattern search at scale
April 14, 2015
Conference Paper
Published in:
IEEE Int. Symp. on Technologies for Homeland Security, 14-16 April 2015.
Topic:
R&D area:
Summary
In recent years, data collection has far outpaced the tools for data analysis in the area of non-traditional GEOINT analysis. Traditional tools are designed to analyze small-scale numerical data, but there are few good interactive tools for processing large amounts of unstructured data such as raw text. In addition to the complexities of data processing, presenting the data in a way that is meaningful to the end user poses another challenge. In our work, we focused on analyzing a corpus of 35,000 news articles and creating an interactive geovisualization tool to reveal patterns to human analysts. Our comprehensive tool, Global Pattern Search at Scale (GPSS), addresses three major problems in data analysis: free text analysis, high volumes of data, and interactive visualization. GPSS uses an Accumulo database for high-volume data storage, and a matrix of word counts and event detection algorithms to process the free text. For visualization, the tool displays an interactive web application to the user, featuring a map overlaid with document clusters and events, search and filtering options, a timeline, and a word cloud. In addition, the GPSS tool can be easily adapted to process and understand other large free-text datasets.
Summary
In recent years, data collection has far outpaced the tools for data analysis in the area of non-traditional GEOINT analysis. Traditional tools are designed to analyze small-scale numerical data, but there are few good interactive tools for processing large amounts of unstructured data such as raw text. In addition to...
READ MORE
Joint audio-visual mining of uncooperatively collected video: FY14 Line-Supported Information, Computation, and Exploitation Program
February 5, 2015
Project Report
Published in:
MIT Lincoln Laboratory Report LSP-116
Summary
The rate at which video is being created and gathered is rapidly accelerating as access to means of production and distribution expand. This rate of increase, however, is greatly outpacing the development of content-based tools to help users sift through this unstructured, multimedia data. The need for such technologies becomes more acute when considering their potential value in critical, media-rich government applications such as Seized Media Analysis, Social Media Forensics, and Foreign Media Monitoring. A fundamental challenge in developing technologies in these application areas is that they are typically in low-resource data domains. Low-resource domains are ones where the lack of ground-truth labels and statistical support prevent the direct application of traditional machine learning approaches. To help bridge this capability gap, the Joint Audio and Visual Mining of Uncooperatively Collected Video ICE Line Program (2236-1301) is developing new technologies for better content-based search, summarization, and browsing of large collections of unstructured, uncooperatively collected multimedia. In particular, this effort seeks to improve capabilities in video understanding by jointly exploiting time aligned audio, visual, and text information, an approach which has been underutilized in both the academic and commercial communities. Exploiting subtle connections between and across multiple modalities in low-resource multimedia data helps enable deeper video understanding, and in some cases provides new capability where it previously didn't exist. This report outlines work done in Fiscal Year 2014 (FY14) by the cross-divisional, interdisciplinary team tasked to meet these objectives. In the following sections, we highlight technologies developed in FY14 to support efficient Query-by-Example, Attribute, Keyword Search and Cross-Media Exploration and Summarization. Additionally, we preview work proposed for Fiscal Year 2015 as well as summarize our external sponsor interactions and publications/presentations.
Summary
The rate at which video is being created and gathered is rapidly accelerating as access to means of production and distribution expand. This rate of increase, however, is greatly outpacing the development of content-based tools to help users sift through this unstructured, multimedia data. The need for such technologies becomes...
READ MORE
NEU_MITLL @ TRECVid 2015: multimedia event detection by pre-trained CNN models
January 1, 2015
Conference Paper
Published in:
TRECVid 2015.
Topic:
R&D area:
Summary
We introduce a framework for multimedia event detection (MED), which was developed for TRECVID 2015 using convolutional neural networks (CNNs) to detect complex events via deterministic models trained on video frame data. We used several well-known CNN models designed to detect objects, scenes, and a combination of both (i.e., Hybrid-CNN). We also experimented with features from different networks fused together in different ways. The best score achieved was by fusing objects and scene detections at the feature-level (i.e., early fusion), resulting in a mean average precision (MAP) of 16.02%. Results showed that our framework is capable of detecting various complex events in videos when there are only a few instances of each within a large video search pool.
Summary
We introduce a framework for multimedia event detection (MED), which was developed for TRECVID 2015 using convolutional neural networks (CNNs) to detect complex events via deterministic models trained on video frame data. We used several well-known CNN models designed to detect objects, scenes, and a combination of both (i.e., Hybrid-CNN)...
READ MORE
Discrimination between singing and speech in real-world audio
December 7, 2014
Conference Paper
Published in:
SLT 2014, IEEE Spoken Language Technology Workshop, 7-10 December 2014.
Topic:
R&D area:
Summary
The performance of a spoken language system suffers when non-speech is incorrectly classified as speech. Singing is particularly difficult to discriminate from speech, since both are natural language. However, singing conveys a melody, whereas speech does not; in particular, a singer's fundamental frequency should not deviate significantly from an underlying sequence of notes, while a speaker's fundamental frequency is freer to deviate about a mean value. The present work presents a novel approach to discrimination between singing and speech that exploits the distribution of such deviations. The melody in singing is typically non known a priori, so the distribution cannot be measured directly. Instead, an approximation to its Fourier transform is proposed that allows the unknown melody to be treated as multiplicative noise. This feature vector is shown to be highly discriminative between speech and singing segments when coupled with a simple maximum likelihood classifier, outperforming prior work on real-world data.
Summary
The performance of a spoken language system suffers when non-speech is incorrectly classified as speech. Singing is particularly difficult to discriminate from speech, since both are natural language. However, singing conveys a melody, whereas speech does not; in particular, a singer's fundamental frequency should not deviate significantly from an underlying...
READ MORE