Razane Tajeddine

Background: Annually, 4% of the global population undergoes non-cardiac surgery, with 30% of those patients having at least one cardiovascular risk factor. It is estimated that the 30-day mortality is between 0.5% and 2%. The main objective of this study is to develop a traditional machine learning (ML) model that provides a cardiovascular risk score for patients older than 50 years undergoing non-cardiac surgery, calculating the risk from the date of surgery until 30 days post surgery, with specific emphasis on interpretability and explainability of the model’s decision-making process. Methods: The NSQIP 2022 dataset was used to build the model. It consisted of a total of 4 97 011 patients after data cleaning. The primary clinical endpoint was death, myocardial infarction, cardiac arrest or stroke at 30 days postoperatively, which occurred in 1.44% of the patients. Different preprocessing techniques were performed for data cleaning and feature selection. The cleaned data were then used to model the selected learning algorithms, including Logistic Regression, Naive Bayes, Random Forest and boosting Decision Tree algorithms (CatBoost, AdaBoost, Light Gradient Boosting Machine (LightGBM, XGBoost, Gradient Boosting). These models were evaluated in terms of the area under the receiver operating characteristic curve (AUROC) and their corresponding 95% CI. Results: For classification, the trained models were evaluated using AUROC on the test set. LightGBM achieved the highest AUROC of 0.9009 with a 95% CI of 0.8889 to 0.9126. The model consisted of six data elements: type of surgery, American Society of Anesthesiology classification, Blood Urea Nitrogen (BUN), sepsis, emergent surgery and mechanical ventilation. Conclusion: In our study, LightGBM classifier proved to be the best model for cardiovascular risk scoring, demonstrating a strong balance between prediction accuracy and generalisation.

The widespread deployment of deep neural networks (DNNs) in critical real-time applications has spurred significant research into their security and robustness. A key vulnerability identified is that DNN decisions can be maliciously altered by introducing carefully crafted noise into the input data, leading to erroneous predictions. This is known as an adversarial attack. In this paper, we propose a novel detection framework leveraging segmentation masks and image segmentation techniques to identify adversarial attacks on DNNs, particularly in the context of autonomous driving systems. Our defense technique considers two levels of adversarial detection. The first level mainly detects adversarial inputs with large perturbations using the U-net model and one-class support vector machine (SVM). The second level of defense proposes a dynamic segmentation algorithm based on the k-means algorithm and a verifier model that controls the final prediction of the input image. To evaluate our approach, we comprehensively compare our method to the state-of-the-art feature squeeze method under a white-box attack, using eleven distinct adversarial attacks across three benchmark and heterogeneous data sets. The experimental results demonstrate the efficacy of our framework, achieving overall detection rates exceeding 96% across all adversarial techniques and data sets studied. It is worth mentioning that our method enhances the detection rates of FGSM and BIM attacks, reaching average detection rates of 95.65% as opposed to 62.63% in feature squeezing across the three data sets.

We present Modular Polynomial (MP) Codes for Secure Distributed Matrix Multiplication (SDMM). The construction is based on the observation that one can decode certain proper subsets of the coefficients of a polynomial with fewer evaluations than is necessary to interpolate the entire polynomial. We also present Generalized Gap Additive Secure Polynomial (GGASP) codes. Both MP and GGASP codes are shown experimentally to perform favorably in terms of recovery threshold when compared to other polynomials codes for SDMM which use the grid partition. Both MP and GGASP codes achieve the recovery threshold of Entangled Polynomial Codes for robustness against stragglers, but MP codes can decode below this recovery threshold depending on the set of worker nodes which fails. The decoding complexity of MP codes is shown to be lower than other approaches in the literature, due to the user not being tasked with interpolating an entire polynomial.

We present Modular Polynomial (MP) Codes for Secure Distributed Matrix Multiplication (SDMM). The construction is based on the observation that one can decode certain proper subsets of the coefficients of a polynomial with fewer evaluations than is necessary to interpolate the entire polynomial. We also present Generalized Gap Additive Secure Polynomial (GGASP) codes. Both MP and GGASP codes are shown experimentally to perform favorably in terms of recovery threshold when compared to other comparable polynomials codes for SDMM which use the grid partition. Both MP and GGASP codes achieve the recovery threshold of Entangled Polynomial Codes for robustness against stragglers, but MP codes can decode below this recovery threshold depending on the set of worker nodes which fails. The decoding complexity of MP codes is shown to be lower than other approaches in the literature, due to the user not being tasked with interpolating an entire polynomial.

In this work, we introduce a differentially private method for generating synthetic data from vertically partitioned data, \emph{i.e.}, where data of the same individuals is distributed across multiple data holders or parties. We present a differentially privacy stochastic gradient descent (DP-SGD) algorithm to train a mixture model over such partitioned data using variational inference. We modify a secure multiparty computation (MPC) framework to combine MPC with differential privacy (DP), in order to use differentially private MPC effectively to learn a probabilistic generative model under DP on such vertically partitioned data. Assuming the mixture components contain no dependencies across different parties, the objective function can be factorized into a sum of products of the contributions calculated by the parties. Finally, MPC is used to compute the aggregate between the different contributions. Moreover, we rigorously define the privacy guarantees with respect to the different players in the system. To demonstrate the accuracy of our method, we run our algorithm on the Adult dataset from the UCI machine learning repository, where we obtain comparable results to the non-partitioned case.

A private information retrieval (PIR) scheme allows a user to retrieve a file from a database without revealing any information on the file being requested. As of now, PIR schemes have been proposed for several kinds of storage systems, including replicated and MDS-coded systems. However, the problem of constructing PIR schemes on regenerating codes has been sparsely considered. A regenerating code is a storage code whose codewords are distributed among nodes, enabling efficient storage of files, as well as low-bandwidth retrieval of files and repair of nodes. Minimum-bandwidth regenerating (MBR) codes define a family of regenerating codes allowing a node repair with optimal bandwidth. Rashmi, Shah, and Kumar obtained a large family of MBR codes using the product-matrix (PM) construction. In this work, a new PIR scheme over PM-MBR codes is designed. The inherent redundancy of the PM structure is used to reduce the download communication complexity of the scheme. A lower bound on the PIR capacity of MBR-coded PIR schemes is derived, showing an interesting storage space vs. PIR rate trade-off compared to existing PIR schemes with the same reconstruction capability. The present scheme also outperforms a recent PM-MBR PIR construction of Dorkson and Ng.

In this paper, the problem of providing privacy to users requesting data over a network from a distributed storage system (DSS) is considered. The DSS, which is considered as the multi-terminal destination of the network from the user’s perspective, is encoded by a maximum rank distance (MRD) code to store the data on these multiple servers. A private information retrieval (PIR) scheme ensures that a user can request a file without revealing any information on which file is being requested to any of the servers. In this paper, a novel PIR scheme is proposed, allowing the user to recover a file from a storage system with low communication cost, while allowing some servers in the system to collude in the quest of revealing the identity of the requested file. The network is modeled as a random linear network, i.e., all nodes of the network forward random (unknown) linear combinations of incoming packets. Both error-free and erroneous random linear networks are considered.

In this paper, the problem of providing privacy to users requesting data over a network from a distributed storage system (DSS) is considered. The DSS, which is considered as the multi-terminal destination of the network from the user's perspective, is encoded by a maximum rank distance (MRD) code to store the data on these multiple servers. A private information retrieval (PIR) scheme ensures that a user can request a file without revealing any information on which file is being requested to any of the servers. In this paper, a novel PIR scheme is proposed, allowing the user to recover a file from a storage system with low communication cost, while allowing some servers in the system to collude in the quest of revealing the identity of the requested file. The network is modeled as a random linear network, i.e., all nodes of the network forward random (unknown) linear combinations of incoming packets. Both error-free and erroneous random linear networks are considered.

The problem of Private Information Retrieval (PIR) from coded storage systems with colluding, Byzantine, and unresponsive servers is considered. An explicit scheme using an [n,k] Reed-Solomon storage code is designed, protecting against t-collusion and handling up to b Byzantine and r unresponsive servers, when n > k+t+2b+r–1. This scheme achieves a PIR rate of n–r–(k+2b+t–1)/n–r . In the case where the capacity is known, namely when k = 1, it is asymptotically capacity-achieving as the number of files grows. Lastly, the scheme is adapted to symmetric PIR.