Hafiz Imtiaz: Previous Projects

Research projects

• Proposed a computationally-light and robust neural network for estimating heart rate in remote healthcare systems. The model can be trained on consumer-grade graphics processing units (GPUs), and can be deployed on edge devices for swift inference. The proposed hybrid model is based on convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) architectures for estimating heart rate from Electrocardiogram (ECG) and Photoplethysmography (PPG) signals. Considering the sensitive nature of the ECG signals, a formal privacy guarantee, differential privacy, is ensured for the model training. A tight accounting of the overall privacy budget of our training algorithm is performed using the Rényi Differential Privacy technique. The model outperforms state-of-the-art networks on a benchmark dataset for both ECG and PPG signals despite having a much smaller number of trainable parameters and, consequently, much smaller training and inference times. The CNN-BiLSTM architecture can also provide excellent heart rate estimation performance even under strict privacy constraints. A prototype Arduino-based data collection system is also developed that is low-cost, efficient, and useful for providing access to modern healthcare services to people living in remote areas. This work led to this Elsevier Healthcare Analytics publication.
• Developed a novel non-negative matrix factorization (NMF) algorithm capable of operating on sensitive data, while closely approximating the results of the non-private algorithm. Additionally, we consider the effect of outliers by specifically modeling them, such that the presence of outliers has very little effect on our estimated differentially-private basis matrix. This theoretical work is also validated on six real datasets of different applications. This work led to this ACM Transactions on Knowledge Discovery from Data publication.
• Developed a novel function computation scheme Gaussian Functional mechanism that ensures strict privacy guaratnee and offers much better utility than existing methods. Additionally, we extend our novel mechanism such that it works seemlessly in decentralized-data settings and propose capeFM, which offers the same privacy-utility trade-off as centralized schemes. This theoretical work is also validated on several real datasets for regression and classification. This work led to this Entropy Information Theory, Probability and Statistics publication.
• Developed an efficient, novel, and implementation-friendly framework for human activity recognition (HAR) using Channel State Information (CSI) of WiFi signal. The framework employs wavelet-based feature extraction and principal component analysis (PCA) based subcarrier fusion to achieve excellent recognition performance of human activity on multiple real datasets, that include variety of environmental conditions. This work led to this Elsevier Digital Signal Processing publication.
• Developed a novel and robust protocol for decentralized differentially private computations that are common in signal processing and machine learning in decentralized data settings. The protocol achieves the same privacy-utility trade-off in the decentralized data setting by employing a correlated noise scheme. We show the effectiveness of the scheme on a decentralized independent component analysis problem on real and synthetic neuroimaging data. This work led to this IEEE Transactions on Signal Processing publication.
• Developed a protocol for decentralized differentially private computations that can achieve the same utility as the pooled-data scenario. The protocol employs a correlated noise scheme and does not require a trusted third party. We analytically show that our framework benefits many machine learning and signal processing problems that appear in practice. This work led to this NeurIPS 2019 workshop publication.
• Developed an algorithm for distributed differentially private canonical correlation analysis (CCA). The algorithm provides a way of computing the CCA subspaces satisfying differential privacy in a distributed setting. The primary achievement of this algorithm is to achieve the same utility as the pooled-data scenario in a distributed setting under the honest-but-curious model. This work led to this publication.
• Developed improved algorithms for distributed differentially private PCA and orthogonal tensor decomposition. These algorithms employ a correlated noise scheme and exploit the "honest-but-curious" network to achieve the same utility as the pooled-data scenario in the distributed setting. This is also the first work for distributed privacy-preserving orthogonal tensor decomposition. This work led to this publication.
• Developed an algorithm for differentially private distributed principal component analysis (PCA). The algorithm provides a way of computing the PCA subspace in a distributed setting, while satisfying differential privacy. PCA subspaces are used in numerous machine learning algorithms as a pre-processing step. This work led to this publication.
• Developed an algorithm for differentially private orthogonal tensor decomposition (OTD). The algorithm provides a way of estimating the parameters of latent variable models. Tensor decomposition for such models has been shown to provide much better accuracy than matrix-based methods. This work led to this publication.
• Developed an algorithm for differentially private canonical correlation analysis (CCA). The algorithm provides a way of computing the CCA subspaces satisfying differential privacy. CCA subspaces can be used to exploit the maximum correlation between different modalities. This work led to this publication.
• Work in progress: exploration of multi-modal data geo-spatial data for correlation analysis and points-of-interests detection while satisfying privacy. A poster based on preliminary work is found here.
• Work in progress: development of algorithms for the collaborative neuroimaging data analysis tool COINSTAC. It provides an easy-to-use platform for computations on distributed datasets and satisfies privacy. A poster based on explanation of this project is found here.
• Developed an open-source Python library dp-stats for commonly used statistics and machine learning algorithms with differential privacy. Included functions: mean, variance, histogram, Principal Component Analysis (PCA), Support Vector Machine (SVM) and Logistic Regression. It also includes examples in iPython notebook format for each function. A poster based on preliminary work and status of the project is found here.
• Developed an algorithm for differentially private distributed joint Independent Component Analysis (djICA). The algorithm provides a way of source separation/matrix factorization in a distributed setting with fMRI data. This work led to this publication.

Course projects

• Phone Number Detection from Dialing Sounds - as a part of the course Digital Signals and Filters.
• STFT Analysis of Bat Sounds - as a part of the course Digital Signals and Filters.
• A Study of Simple and Practical Algorithm for Sparse Fourier Transform - final course project for the course Digital Signals and Filters.
• Analysis of the Performance of a Bank Queue System using Markov Chain - final course project for the course Stochastic Signals and Systems.
• Empirical Comparison of Classification Performance of Differentially-private Principal Component Analysis Algorithms Using Support Vector Machine - final course project for the course Convex Optimization.
• Implementation and Empirical Comparison of Four Face Recognition Algorithms - as a part of the course Advanced Topics in DSP - Biometrics.
• Empirical Comparison of Sparse Embedding and K-SVD - final course project for the course Advanced Topics in DSP - Biometrics.
• Empirical Comparison of Tensor and Matrix based Methods for Image Classification - final course project for the course Image Coding and Processing.