Secure Signal Processing

By Shantanu Rane and Mauro Barni

Over the past decade, individual users and businesses have entrusted increasing amounts of data to online storage and computing facilities. The ease in sharing and processing of this data has fueled an unprecedented rise in e-commerce, which is set to increase even further with the advent of cloud computing. At the same time, it has become necessary to develop techniques that allow sophisticated data processing while preserving the privacy of individual users and intellectual property of businesses.

A powerful way to address security and privacy concerns is to perform computation and processing on signals that have been obfuscated either by encryption or secret sharing or other privacy-preserving primitives. Mathematical manipulation of signals that are encrypted or otherwise hidden, is a fascinating challenge. In response to this emphasis on security and privacy, there has been a surge of interest in fundamental and applied aspects of secure signal processing, fueled by researchers from the cryptography, signal processing, data mining communities.

Recognizing the relevance of this area and the recent increase in research interest from various communities, the Information Forensics and Security Technical Committee of the IEEE Signal Processing Society organized a special session on Secure Signal Processing at the recently held ICASSP 2011 at Prague on May 25, 2011.

The goals of the special session were threefold: (a) To highlight, by means of examples, the recent enabling technologies in secure signal processing, (b) To draw the attention of the broader ICASSP community to this fertile area of truly interdisciplinary research, and (c) To give sense of what signal processing is currently possible with secure primitives, with an eye on the key unsolved issues in the field.

The special session brought to the conference six papers on various aspects of secure multiparty computation, with a view to acquaint the signal processing community with state-of-the-art techniques that allow limited signal processing on encrypted information, and to highlight the challenges that lie ahead, if realtime secure signal processing is to become a reality.

In a talk titled "Is Secure Multiparty Computation Any Good in Practice?", Dr. Claudio Orlandi gave a computer science-oriented view of the problem, gave formal definitions for what it means for a computation protocol to be secure and introduced homomorphic encryption techniques which allow computation to be performed directly in the encrypted domain. Following this Dr. Tiziano Bianchi presented an information-theoretic view of secure computation, and analyzed additive and multiplicative blinding techniques, which are ubiquitous in secret sharing protocols.

Prof. Min Wu then addressed several open problems in secure video processing. Recognizing the increasing volume of video data, this talk described the possibility of deriving low-dimensionality representations of "fingerprints" of multimedia data for use in collusion-resistant copyrighting. The talk also highlighted the fact that multimedia data is of unequal importance and so bits of unequal importance may benefit from different protection techniques.

There were three more talks which described different applications of homomorphic functions, a class of functions that allows operations such as additions and multiplications in the encrypted domain. Dr.
Zekeriya Erkin described how a privacy-preserving recommendation system can be built using additive homomorphisms. This allows a movie rental company, for example, to securely solicit movie ratings from its users and generate recommendations for its customer base, without any knowledge of the movies watched by individual customers. Dr. Juan Ramon Trancoso-Pastoriza presented a technique for perfoming adaptive filtering using the properties of homomorphic functions. Dr. Shantanu Rane described how additively homomorphic functions can be used to perform inference on private data using statistical models in a client- cloud framework. In particular, this talk showed how a client can send encrypted speech data to a server which then performs oblivious keyword spotting using trained Hidden Markov Models in the cloud.