Privacy Preserving Computation


This work involves inferring knowledge from data in such a way that either the data, or the inference, or both, have to be protected from adversaries. The approaches followed depend upon the kinds of security and privacy guarantees desired and the strength of the adversary. Primarily, we employ computationally secure primitives, such as homomorphic encryption, for manipulating data in the encrypted domain. We also apply information-theoretic approaches such as secret sharing, and statistically private mechanisms such a differential privacy.

The underlying theme is to figure out ways in which data can be secured not only during storage and transmission as is usually done, but also during the process of computation. Our current challenge is to design efficient, privacy-aware protocols that perform expressive analytics and machine learning operations. We tackle interesting problems in the realm of enterprise as well as personal privacy.

Some representative publications:

  1. S. Rane and P. Boufounos, Privacy-Preserving Nearest Neighbor Methods, IEEE Signal Processing Magazine, Vol. 30, No. 2, pp. 18--28, March 2013. [link] [preprint
  2. S. Rane, J. Freudiger, A. Brito, E. Uzun, Privacy, Efficiency and Fault Tolerance in Aggregate Computations on Massive Star Networks, IEEE Workshop on Information Forensics and Security (WIFS 2015), Rome, Italy, November 2015. [paper]

Visual Inference via Randomized Embeddings


This research is concerned with efficient machine learning operations on high-dimensional signals, such as images and videos. The philosophy is to map high-dimensional signals into a low-dimensional vector space such that some aspects of the signal geometry are preserved without necessarily targetting faithful recovery of the signals themselves.

Most of the work that we have done so far has centered around distance-preserving embeddings based on the Johnson-Lindenstrauss Lemma. Particularly, to communicate the signal embeddings from a client to a database server, it is necessary to first quantize them. We have extended the Johnson-Lindenstrauss Lemma to quantized embeddings. Furthermore, by carefully designing the quantizer, some embeddings can be endowed with an interesting information-theoretic privacy property. This makes them useful in the design of privacy-aware nearest neighbor protocols.

Some representative publications:

  1. M. Li, S. Rane and P. Boufounos, Quantized Embeddings of Scale-Invariant Image Features for Mobile Augmented Reality, IEEE International Workshop on Multimedia Signal Processing (MMSP 2012), Banff, Canada, September 2012. [paper
  2. P. Boufounos and S. Rane, Secure Binary Embeddings for Privacy-Preserving Nearest Neighbors, IEEE International Workshop on Information Forensics and Security (WIFS 2011), Iguazu Falls, Brazil, November 2011. [paper]

Biometric Template Protection


This work encompasses the theory, software implementation, and standardization of secure fingerprint-based biometric recognition systems. As a biometric is an irreplaceable attribute of a person, loss or theft of the biometric leaves her vulnerable to spoofing, identity theft and loss of sensitive data. To insulate a user from such vulnerabilities, we construct biometric recognition methods that do not require the system to store a person's biometric (or biometric features) in the clear.

This work is interdisciplinary. It involves the use of image processing and machine learning to identify and extract discriminative features from human biometrics. To prevent the extracted features from being accessed by adversaries while also allowing the comparisons of test biometrics against the enrolled biometric identities, we have employed information theoretic primitives, such as fuzzy vaults, and cryptographic primitives, such as homomorphic encryption.

Some representative publications:

  1. S. Rane, Y. Wang, S. C. Draper and P. Ishwar, Secure Biometrics: Concepts, Authentication Architectures, and Challenges, IEEE Signal Processing Magazine, Vol. 30, No. 5, pp. 51--64, September 2013. [link] [preprint]
  2. Y. Wang, S. Rane, S. C. Draper and P. Ishwar, A Theoretical Analysis of Authentication, Privacy and Reusability Across Secure Biometric Systems, IEEE Trans. Information Forensics and Security, Vol. 7, No. 6, pp. 1825--1840, December 2012. [arXiv]

Distributed Video and Image Coding

The goal of this work was to develop distributed compression frameworks for multi spectral and hyper spectral imagery. In contrast to traditional compression architectures, this approach was developed for applications in which the encoding has to be performed under strict constraints on computational complexity, available storage and battery life. In return, a more complex decoder can be tolerated. These situations occur, for example, in sensor networks and satellite-based data compression. 

This research drew on information theoretic results on source coding with side information. Primarily these included the lossless distributed compression problem studied by Slepian and Wolf and its rate-distortion counterpart studied by Wyner and Ziv. It also leveraged results from the more recent research activity in distributed video coding.

Some representative publications:

  1. Y. Wang, S. Rane, P. Boufounos and A. Vetro, Distributed Compression of Zerotrees of Wavelet Coefficients, IEEE International Conference on Image Processing (ICIP 2011), Brussels, Belgium, October 2011. [paper]
  2. B. Girod, A. Aaron, S. Rane, D. Rebollo-Monedero, Distributed Video Coding, Proceedings of the IEEE, Special Issue on Advances in Video Coding and Delivery, Invited Paper, Vol. 93, No. 1, pp. 71-83, January 2005. [paper]