Abstract

Speaker: Diana A. Vasile

Title: End-to-end encryption for software-as-a-service

Abstract: Cloud and software-as-a-service applications such as Google Docs, Evernote, iCloud and Dropbox are very convenient for users, but problematic from a security point ofview. As these services process data in unencrypted form on their servers, users must blindly trust the cloud provider to prevent unauthorised access and to maintain integrity of the data. A security breach of the cloud provider could have disastrous consequences.

In this project, we are exploring techniques for Trust-Reducing Verifiable Exchange of data (TRVE Data, pronounced “true data”). Our goal is to create the foundations for applications that are as usable and convenient as today’s cloud services, while reducing the amount of trust that is placed in third parties.

In particular, we are exploring cryptographic techniques for improving confidentiality, preventing unauthorised access to sensitive data; integrity, ensuring data items cannot be tampered with; and availability, ensuring continued access to data in the face of malice.

This project touches on many areas of computer security, distributed systems and human-computer interaction. We are applying end-to-end encryption and integrity proof techniques to the domain of databases and real-time collaborative applications. We are designing user interfaces to encourage safe user behaviour. We are building upon a long history of distributed systems research to create data synchronisation mechanisms that are robust to both malicious and accidental network interference.

——————

Speaker: Naruemon Pratanwanich (Ploy)

Title: Predicting disease-therapeutic gene targets from multiple evidences

Abstract: Accurate prediction of disease-gene associations requires the integration of different evidences, such as genome-wide association studies, literate databases and gene expression experiments. Frequently, these data are complementary and capture different aspects of the landscape of gene-disease associations. A certain degree of consensus among these evidences makes us more confident on true associations.

Here, our objective is to impute the missing associations between drug targets and diseases as well as discover the agreement over multiple datasets (views). Considering multiple datasets simultaneously allows those diseases/genes that are not studied at all in one view can be imputed by borrowing the information from other views.

In order to address this issue, we develop a joint matrix factorisation approach that shares the same latent factors across different views, where each latent factor is considered a soft-cluster for both diseases and targets. Still, we allow each view to be mapped from the latent space to their own data space with different transformation functions. In particular, our model automatically chooses the transformation parameters e.g. scaling and location shifting for each view such that transformed data are well-modelled by the joint Gaussian-likelihood matrix factorisation. We believe that including the transformation as an integral part of the matrix factorisation model rather than as an ad-hoc step not only make the multi-view analysis possible, but also yields the better performance on prediction.

We apply the approach to disease-gene matrices generated by the Centre for Therapeutic Target Validation and present results on how this model-based approach can help to impute missing relationships and unravel core consensus across different views.