2023
Authors
Brito, CV; Ferreira, PG; Portela, BL; Oliveira, RC; Paulo, JT;
Publication
IEEE ACCESS
Abstract
The adoption of third-party machine learning (ML) cloud services is highly dependent on the security guarantees and the performance penalty they incur on workloads for model training and inference. This paper explores security/performance trade-offs for the distributed Apache Spark framework and its ML library. Concretely, we build upon a key insight: in specific deployment settings, one can reveal carefully chosen non-sensitive operations (e.g. statistical calculations). This allows us to considerably improve the performance of privacy-preserving solutions without exposing the protocol to pervasive ML attacks. In more detail, we propose Soteria, a system for distributed privacy-preserving ML that leverages Trusted Execution Environments (e.g. Intel SGX) to run computations over sensitive information in isolated containers (enclaves). Unlike previous work, where all ML-related computation is performed at trusted enclaves, we introduce a hybrid scheme, combining computation done inside and outside these enclaves. The experimental evaluation validates that our approach reduces the runtime of ML algorithms by up to 41% when compared to previous related work. Our protocol is accompanied by a security proof and a discussion regarding resilience against a wide spectrum of ML attacks.
2022
Authors
Baptista, D; Ferreira, PG; Rocha, M;
Publication
Abstract
2023
Authors
Moraes, A; Moreno, M; Ribeiro, R; Ferreira, G;
Publication
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Abstract
The accurate prediction of biological age can bring important benefits in promoting therapeutic and behavioural strategies for healthy aging. We propose the development of age prediction models using multi-modal datasets, including transcriptomics, methylation and histological images from lung tissue samples of 793 human donors. From a technical point of view this is a challenging problem since not all donors are covered by the same data modalities and the datasets have a very high feature dimensionality with a relatively smaller number of samples. To fairly compare performance across different data types, we’ve created a test set including donors represented in each modality. Given the unique characteristics of the data distribution, we developed gradient boosting tree and convolutional neural network models for each dataset. The performance of the models can be affected by several covariates, including smoking history, and, most importantly, by a skewed distribution of age. Data-centric approaches, including feature engineering, feature selection, data stratification and resampling, proved fundamental in building models that were optimally adapted for each data modality, resulting in significant improvements in model performance for imbalanced regression. The models were then applied to the test set independently, and later combined into a multi-modal ensemble through a voting strategy, predicting age with a median absolute error of 4 years. Even if prediction accuracy remains a challenge, in this work we provide insights to address the difficulties of multi-modal data integration and imbalanced data prediction. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
The access to the final selection minute is only available to applicants.
Please check the confirmation e-mail of your application to obtain the access code.