I am working as part of the Undergraduate Research Opportunities Program (UROP) at the Information Processing and Communications Lab at Imperial College London.

Privacy-Aware Sharing in Smart Grids (canceled)

In the new generation electricity supply networks, smart meters play a key role in recording the electricity consumption of a household by reporting it to the utility provider in real-time. However, this real-time information also carries sensitive personal information about the user’s activities. Therefore, sending a distorted version of the smart meter readings that satisfies the consumer’s demand has been considered. The aim is to minimize the mutual information between the consumer’s data and its distorted version.

At the beginning of the UROP, I have been reading about the privacy-aware data sharing in the context of the new generation electricity supply network, Smart Grids, using reinforcement learning. I have worked on extending the Deep-RL work discussed in the following paper from the single-user case to multi-user cases using adversarial Reinforcement Learning. While formalizing the context of this problem, the project canceled because it turned out to be unrealistic.

Adversarial Privacy for Activity Monitoring

With the advance of smartphones, wearables devices, and other Internet of Things devices, there is a variety of sensors that are generating time-series measurements of your daily activity. These sensors allow the emergence of new services that are beneficial to several areas: health monitoring, safety, and productivity. While cameras, microphones, and the location are perceived as privacy-sensitive, the privacy implication of activity recognition sensors is still underestimated. The detailed time-series user data shared with untrusted third parties could be used to infer private and sensitive information user information (i.e. if the user is smoking).

At the end of this UROP, I have formulated the context of the problem, investigated the different adversarial learning architectures that could to extend work done in Replacement Autoencoders to have it learn the features to hide without any human interaction. Also, I have preprocessed, analyzed, and visualized the available data. Eventually, I have made the first basic model to work using the Generative Adversarial Network (GAN) approach.

This project is currently being expanded as part of my Final Year Project

• GAN
• machine learning
• reinforcment learning