M Tech thesis projects supervised:
(1) Privacy preserving face detection and its application to face recognition based on composite masks, 2015
(2) Generating and Identifying patterns in Shuffles, 2015
(3) Detecting rotation and Interpolation operations in images, 2014
(4) Privacy preserving face biometric comparison, 2014
(5) Face acquisition in dynamic environments, 2014
(6) Face retrieval based on Eigen-correspondences, 2013
(7) Audio fingerprinting for multi-lingual retrieval of Indian film songs, 2013
(8) Estimation of light source direction for blind forensics of images using vanishing points, 2013
(9) Anonymous Computing Protocols for E-voting and E-tendering, 2012
(10) A Code and Domain Independent Traitor Tracing System, 2012
(11) A Signal Processing approach to Authenticating Encrypted Images, 2011.
(12) Generating Non-invertible and Cancellable Hashes from Fingerprints for Secure Authentication, 2011.
(13) Parametric transforms for Image Watermarking, 2011.
(14) Compression sensitive semi-fragile watermarking for image authentication, 2010.
(15) Fractal based image authentication, 2010.
(16) Visual cryptography with fingerprinting, 2009.
(17) Design and analysis of H.264 video encryption schemes for real time video applications, 2009.
(18) Biometric Authentication: Principles of template protection and Gabor filter based feature extraction, 2009.
Selective abstracts of MTech theses:
Privacy preserving computation
M. Tech thesis of Avik Hati (Jan 2011 to May 2012)
With the rapid development in information technology and the correspondingly increasing threats in information security, privacy preserving computation has found many practical applications such as electronic voting, electronic tendering, access control system etc. In this context, a multi-authority Electronic Voting Protocol and an Electronic Tendering protocol have been proposed independently in this thesis. The main aim in designing these systems is to keep the users, of whatever the system may be, anonymous to everyone else including other users, different authorities involved and the manager of the system. Simultaneously, the systems also maintain other important characteristics of security and are robust to different attacks like collusion.
As the first contribution, an Electronic Voting Protocol has been developed, analyzed and improved. Here, vote generation is done by addition of keys (secret integers) and votes which are represented by a set of integers previously fixed and the vote counting is done using a pre-computed table. The system has the flexibility to discard the encryption of votes if there is no threat of active attack during message transmissions. A novel technique is used in the vote counting phase by expressing a number as the linear combination of voting choices to reduce the complexity associated with the large sized pre-computed table used in Kobayashi's e-Voting proposition. Next, a modified version of one-way homomorphic encryption function has been applied to all the steps in the original algorithm to stop the voters from modifying messages to change the result of election in favor of a particular candidate, by performing all the calculations in the encrypted domain. Finally, to reduce the effect of large-valued voting choices and to detect manipulations in each received voting message, key vectors have been used for vote generation.
As the second contribution, an Electronic Tendering protocol has been developed based on trusted third party architecture. The focus in this protocol has mainly been on finding only the tenderer with the minimum amount and only the corresponding minimum amount whereas other tenderers' amounts are not disclosed.
A Code and Domain Independent Traitor Tracing System
M. Tech thesis of Gnana Praveen (Jan 2011 to May 2012)
When multimedia content is distributed over a broadcast network, the content is usually encrypted in such a way that only the legitimate users can be able to access the content by decrypting it's encrypted version using a certain set of keys. Once they get access to the content they may be willing to disclose their data by constructing a pirated version which allows illegitimate users to access the content. Hence it is desirable to trace the users who involved in creating the pirated version. Traitor Tracing schemes were introduced to address this problem. Conventional traitor tracing approaches focus on creating associations between multimedia fingerprints through anti-collusion codes. Most conventional tracing schemes are symmetric in nature, where partial knowledge of the embedding process such as domain and embedding function is required at the time of fingerprint detection and tracing. In most piracy lawsuits the handling of evidence and extraction of fingerprint information is done by an authority other than the source (or fingerprint embedder). To prevent tampering of fingerprint information and evidence by the tracing authority itself there is a need for privacy preserving tracing mechanism. Such a mechanism naturally demands a decoupling of the embedding and detection processes. In this thesis, we explore the possibility of creating a code and embedding domain independent fingerprint detection strategy by performing an eigen-decomposition of the fingerprinted images. The resultant Basis which spans the space of valid fingerprinted images and their respective projections are used as side-information for traitor tracing. Any colluded image is projected onto this eigen-space and the nearest neighbors are identified as potential traitors. But the idea of deriving a basis tailored to the spread of data (fingerprints) is not entirely secure. Hence, we have enhanced the privacy of the proposed architecture by investigating the scope of random projections in designing a privacy preserving detection mechanism. Random Projections are linear transformations known to preserve the distances in lower dimensional spaces. Furthermore the spread of the data (fingerprinted images) along these random atoms is almost uniform, devoid of any form of bias towards a certain atom. Such a linear transformation can be made one way analogous to a distance preserving robust hash of the fingerprinted image. This mechanism allows pirated images to be projected onto this randomized basis and comparison along with traitor tracing is done in the projected space, making the detection mechanism independent of the fingerprint concealment methodology.
Generating Non-invertible and Cancelable Hashes from Fingerprints for Secure Authentication
M. Tech thesis of Priyanka Das (Jan 2010 to May 2011)
Considering the privacy and security issue, an alignment-free novel fingerprint hashing algorithm is proposed, which uses a graph comprising of the inter-minutia minimum distance vectors originating from the core point as a feature set called the Minimum Distance Graph (MDG). Extracted features are rotation and translation invariant, also robust to noise caused by minutiae insertions and deletions. Comparison between hashes has been implemented using a Corresponding Search Algorithm (CSA). Based on experiments conducted on the FVC2002-DB1a and FVC2002-DB2a databases, we obtain an Equal Error Rate of 3:63%. Our algorithm reduces the computational cost associated with the fingerprint matching in terms of the hash comparisons while, balancing the false alarm rate. A preliminary analysis of the security strength shows that the invertibility of the hash is intractable to a brute-force attack. An extension to generate cancelable template from the same set of features and MDG hash is also discussed.
A Signal Processing Approach to Authenticating Encrypted Images
M. Tech thesis of Sachin Kashyap (Jan 2010 to May 2011)
Every form of data either generated naturally or artificially has a unique statistical signature. The process of encryption is designed specifically to mask (or destroy) this signature (for most applications). By integrating the process of source/transform coding with encryption, one can allow a portion of this signature to surface. When this encoded/encrypted stream is supplemented with suitable and carefully designed public information, a third party can be made to attest the origin and integrity of the data stream without being able to tap into the contents. A transposition cipher (T) primarily alters the positions of the plaintext according to some secret key but tends to preserve certain statistical characteristics of the underlying plaintext. When applied towards Image encryption, we observed that the first and second order block statistics are invariant to transposition. This allows us to construct execute the hashing operation in the encrypted domain, opening the doors to privacy preserving data authentication.
Parametric Transforms for Digital Watermarking
M. Tech thesis of Suresh Nakum (Jan 2010 to May 2011)
Every signal space has a basis or a set of bases which permit an orthogonal or non-orthogonal decomposition of the signals into a set of components aligned along the basis vectors. The choice of basis is very often tailored towards the characteristics of the signal and designed to provide compact support. Some of the standard transforms (or basis sets) applied to discrete signals are the discrete Fourier transform (DFT), discrete Hadamard transform (DHT) and discrete cosine transform (DCT). Most watermarking techniques do not conceal the transform domain in which the watermark signal is embedded. Theoretically there are many possible choices of bases which span a given signal space. However only a selective few form an orthogonal basis. If orthogonality is sacrificed and the choice of basis is made arbitrary, there are seemingly infinite possible decompositions of a particular signal. One way to construct a family of basis functions is by parameterizing the standard transforms such as DCT, DFT and DHT. The set of parameters which are used to identify a particular basis within this family form a secret key. Thus through this secret basis one can construct a hidden domain in which a watermark can be concealed. In our work explore the parameterization of GDFT and GDCT kernels and apply these constructions to areas such as Digital Fingerprinting, Secure Image Hashing and Multiple Watermarking.
The search for self-similarity: Application of Fractals to image compression and watermarking
M. Tech thesis of Balaram. A. Reddy (Jan 2009 to June 2010)
It is natural to expect highly complex, yet tightly inter-connected systems such as social grids, web traffic or cellular growth to exude this form of self-similar behaviour wherein even the tiniest fraction can be chosen to represent the whole. In much simpler systems, the search for self-similarity (statistical) is simply an excuse to achieve one of the following: (1) Signal classification, (2) Authentication of information, (3) Compression. In this work we exploit the self-similarity in natural images for creating associations across image blocks. This self-similarity map is used for verifying if certain parts of the original image have been manipulated, as the latter would break the association between two previously declared self-similar blocks. The intermediate "key" is the set of affine transformations derived by partitioning and matching domain and range blocks from the original image which are kept private, but the partitioning and matching algorithms are made public. Only the block-map is preserved as the "self-similar hash". We found that this algorithm is robust to several content preserving operations and captures the diversity across a variety of natural images.
Creating audit trails: Application of Semi-fragile watermarks
M. Tech thesis of Durgaprasad. Mandala (Jan 2009 to June 2010)
The flight of an albatross is likely to leave a trail of waning, but characteristic pressure waves. In a similar fashion one can attach a life span to the authenticity of an electronic document circulated amongst a closed group of individuals. Every event, defined by the union of the following two operations: (a) decoding/decompressing/viewing and (b) re-compressing/storing can be associated with the controlled fading of a time-stamped semi-fragile watermark, embedded in the document.
As a first step in this direction, we begin investigating the use of semi-fragile watermarks for detecting intentional manipulations to images. Intentional attacks include modification to selective portions of the content, cropping, localized blurring and geometric operations such as scaling and rotation. The watermark is embedded in specific coefficients in the DCT domain by changing the least significant bits (LSBs) of DCT transform coefficients. The image is partitioned into blocks of size two-by-two and then exactly one coefficient amongst four coefficients is randomly chosen for embedding. Retrieval of the watermark can be done in a semi-blind fashion with the help of only a key containing the positions of the embedded coefficients and some additional side information regarding the original un-watermarked image.