Privacy Aware Multimedia Computation

Multidimensional streaming signals like audio, videos, and multi-spectral imagery are truly the biggest of the Big Data. Collected by sensors at the network edge, these data often contain highly sensitive information and their misuse can lead to significant invasion of privacy. For example, networks of surveillance cameras have been used by some countries in tracking their citizens, and pervasive use of IoT devices in smart homes opens door to recording private behaviors and interactions. It is highly challenging to apply traditional privacy enhancing technologies to handle these data. Cryptographic techniques like homomorphic encryption and garbled circuits are too computationally intensive for any practical applications. The more efficient differential private schemes, on the other hand, rely on additive noise and may not be able to provide adequate protection on semantic contents.

In this project cluster, we have developed a number of techniques to process multidimensional signals while ensuring their privacy. First, I will talk about how to exploit the nature of the signals in speeding up cryptographic computation. Specifically, we have developed a novel optimized design of a garble-circuit based iris-code matching algorithm where we can achieve significant speedup using a common iris mask. Second, we have investigated the use of secret sharing as a faster and more convenient privacy platform for distributed signal computation. Secret sharing does not require key distribution and can achieve information theoretic security without long integer fields. However, secret sharing is prone to collusion attacks. We have proposed a peer-to-peer computational framework that combines cryptographic and game-theoretic countermeasures to prevent such attacks. Finally, we are currently studying generative adversarial network (GAN) and random neural networks to protect signal privacy in distributed deep learning. We use GAN to learn the statistics of sensitive facial data at local sites and generate privacy-preserving synthetic data for public centralized learning. While GAN at network edge can be computationally intensive, random neural network offers a much lighter weight privacy-enhancing protocol suitable for a broad range of IoT applications.

Selected Publications: