Efficient Privacy Preserving Protocols for Visual Computation Maneesh Upmanyu IIIT Hyderabad Advisors: C. V. Jawahar , Anoop M. Namboodiri, Kannan Srinathan, Center for Visual Information Technology Center for Security, Theory & Algorithmic Research IIIT- Hyderabad Security and Privacy of Visual Data Broad Objective • Development of secure computational algorithms in computer vision and related areas. – To develop “highly-secure” solutions – To develop “computationally efficient” solutions IIIT Hyderabad – To develop solutions to problems with immediate impact Project Web-Page: http://cvit.iiit.ac.in/projects/SecureVision Research Directions Private Content Based Image Retrieval (PCBIR) Blind Authentication: A Secure Crypto-Biometric Verification Protocol Efficient Privacy Preserving Video Surveillance Feature vector (fquery) Root Info fquery, f(A1) Q1 A1 fquery, f(A2) Q2 A2 …….. Publication: Maneesh Upmanyu, Anoop M. Namboodiri, K. Srinathan and C.V. Jawahar; Efficient Privacy Preserving Video Surveillance: Proceedings of the 12th International Conference on Computer Vision (ICCV 2009) IIIT Hyderabad Publication: Maneesh Upmanyu, Anoop M. Namboodiri, K. Srinathan and C.V. Jawahar; Blind Authentication - A Secure Crypto-Biometric Verification Protocol: Appears in IEEETransactions on Information Forensics and Security (IEEE-TIFS), June 2010 Publication: Shashank J, Kowshik P, Kannan Srinathan and C.V. Jawahar; Private Content Based Image Retrieval; In Proceedings of Computer Vision and Pattern Recognition (CVPR 2008) Our Security Goal • What is meant by ‘Privacy’? – Design protocols to limit the information leakage through what is learned in addition to the designated output. • What is the ‘Adversary Model’? – Semi-honest vs. Malicious adversary • Analysis outline: IIIT Hyderabad – Correctness – Security – Complexity Assumptions • Reliable and secure communication channel • Players are passively corrupt, that is, honest but curious. • Players are computationally bounded. IIIT Hyderabad • Players do not collude. Thesis Objective • Traditional Approaches uses highly interactive protocols. – Limitation: massive datasets – Example: Blind Vision • Paradigm Shift – Compute directly in encrypted domain. • Encrypt -> Communicate -> Compute -> Decrypt – Domain specific encryption schemes. • PKC is data independent and generic. IIIT Hyderabad – Can the paradigm be generic yet efficient? Contribution of Thesis IIIT Hyderabad A method that provides provable security, while allowing efficient computations for generic vision algorithms have remained elusive. We show that, one can exploit certain properties inherent to visual data to break this seemingly impenetrable barrier. IIIT Hyderabad Dilemma of Privacy vs. Accuracy What is Blind Authentication? A biometric authentication protocol that does not reveal any: – information about the biometric samples to the authenticating server. IIIT Hyderabad – information regarding the classifier, employed by the server, to the user or client IIIT Hyderabad Biometric Authentication System Primary Concerns in a Biometric System • Template Protection • Non-Repudiable • Network and Client-side Security IIIT Hyderabad • Revocability IIIT Hyderabad Previous Work “A template protection scheme with provable security and acceptable recognition performance has thus far remained elusive.” – A.K. Jain, Eurasip 2008 Homomorphic Encryption • An encryption scheme using which some algebric operation , like addition or multiplication, can be directly done on the cipher text. Let x1 = 20 and x2 = 22, to compute x1+x2 = 42 Use an encryption scheme, for example E(x) = ex Server stores E(x1) = e20 and E(x2) = e22 IIIT Hyderabad Compute using encrypted data y = E(x1) E(x2) = e20.e22 = e42 Decrypt z = D(y) = ln(y) z = D(y) ln (e42) = 42 User Enrollment IIIT Hyderabad Enrollment based on a trusted third party. IIIT Hyderabad Authentication using a Linear Kernel Extensions to Kernels & Neural Networks • Kernel based classifier uses a discriminating function like • Similarly, in Neural Network the basic units are for example perceptron or sigmoid • Model above functions as arithmetic circuits consisting of add and multiplication gates over a finite domain. IIIT Hyderabad • Consider two encryptions E+ and E* Implementation and Analysis • Experiments designed to evaluate the efficiency and accuracy of proposed approach. • For evaluation, an SVM based verifier based on clientserver architecture was implemented. – Accuracy: as no assumptions are made, accuracy remains same. IIIT Hyderabad • Verified this on various public domain (UCI, Statlog) datasets. IIIT Hyderabad Case study shows that matching using fixed length feature representation is comparable to variable length methods such as dynamic warping. Security, Privacy and Trust • Server Security – Template database security – Hacker sitting in server • Client Security – Hacker has user’s key or biometric – Passive attacks at client end IIIT Hyderabad • Network Security – Network is susceptible to snooping attacks Advantages of Blind Authentication • Fast and Provably Secure authentication without trading off accuracy. • Supports generic classifiers such as Neural Network and SVMs. • Useful with wide variety of fixed-length biometric-traits. IIIT Hyderabad • Ideal for applications such as biometric ATMs, login from public terminals. Proposed Surveillance System Plain Video Encrypted Video Processed Video Result Video Captured by Camera As seen by one of the Computational Servers As seen by the Computational Server Received by Observer IIIT Hyderabad How do we carry out surveillance on ‘Randomized’ images ? Motivation Can we do surveillance without ‘seeing’ the original video ? Ability to run video surveillance algorithms, completely in encrypted domain can address most privacy concerns. IIIT Hyderabad Existing methods are either too slow for surveillance applications or do not provide provable privacy. Paradigm Shift Trusted Third Party Selective Encryption (TTP) (Smart Camera) In practice, do not have the luxury of a trusted entity Homomorphic Encryption (Doubly) IIIT Hyderabad Computationally expensive No provable privacy, costly and tedious to upgrade Traditionally Explored Paradigms Secure Multiparty Computation (SMC) Highly inefficient, High level of privacy, an overkill in practice We use the paradigm of secret sharing to achieve private and efficient surveillance. Protocol in a nutshell IIIT Hyderabad Propose a ‘Cloud-Computing’ based solution using k>2 non-colluding servers Shatter Merge Compute Image Result • The camera splits each captured frame F, into k ( > 2 ) using pixel level shatter function: •shares carry out aaof basic operation f on the input eachby •ToThe results operations on the shares areimage, integrated server blindly carries the equivalent basic operation f’ on the observer using aout merge function ( CRT), to obtain final itsresult. share. • Each share is then sent to an independent server for processing. Secret Sharing • A method of distributing a secret among a group of servers, such that: IIIT Hyderabad – Each server on its own has no meaningful information – Secret is reconstructed only when all shares combine together • Existing methods are highly inefficient • Asmuth-Bloom overcomes this limitation by working in Residue Number System (RNS). Example to do Addition in RNS RNS ( m1 = 37, m2 = 49; M = m1 x m2 = 1813) X = 973%(m1, m2) (x1, x2) = (11, 42) Y = 678%(m1, m2) (y1, y2) = (12, 41) Shatter: f(x) = (x.S+h) mod mi x1 = 11, y1 = 12 x2 = 42, y2 = 41 z1 = (x1 + y1) % m1 = (11+12) % 37 = 23 z2 = (x2 + y2) % m2 = (42+41) % 49 = 34 IIIT Hyderabad Merge: m(xi, mi) = CRT(xi, mi) /S CRT (z1, z2) Z = 1651 Data Properties • While general purpose secure computation appears inherently complex and oftentimes impractical. – We show certain properties of the data can be used to ensure efficiency while ensuring privacy. • Following properties are of interest to us. IIIT Hyderabad – – – – Limited and Fixed Range Scale Invariant Approximate Nature Non-General Operands Characteristics of the System IIIT Hyderabad Preserve Privacy • Carry out surveillance on random looking images. Light weight • Encrypted domain representation should allow efficient computations. Limited data expansion • Obfuscation process should not blow up the video data. Secure Storage • Obfuscation should be provably secure to ensure security at un-trusted servers. Reconstruction of data • Only authorized people should be able to recover original plain video. Implementation Challenges • Representation of negative numbers: Use an Implicit sign representation. – Use (0, M/2) as positive and rest as negative. – Sign conversion is carried out using additive inversion of Z. • Overflow and Underflow: Operations are valid and correct as long as range of data is (-M/2, M/2). IIIT Hyderabad • Integer Division and Thresholding: RNS domain is finite and hence not all divisions are defined. – Dividing integer A by B is defined as A/B = (ai.bi-1) mod mi • Defining Equivalent operations: For every f(x), we need to define f`(x) such that merging f`(xi) would give f(x). IIIT Hyderabad Experimental Results IIIT Hyderabad Properties of the Protocol • Servers are un-trusted and the network may be insecure. • Near loss-less (PSNR~51). data encoding • No compromise in accuracy. • Inexpensive capture device, and a unidirectional data flow. • Negligible overheads to make private computation practical. Circumvent theoretical bounds. Extremely efficient over SMC Not only efficient, but also provably secure Scalable, inexpensive and generic, thus practical IIIT Hyderabad • Secure as long as servers do not collude. Our approach shows that privacy and efficiency co-exists in the domain of visual data K-Means Clustering IIIT Hyderabad • Data clustering is one of the most important techniques for discovery of patterns in a dataset. • K-Means clustering is a simple and extensively used technique that automatically partitions a dataset into k clusters. • The technique becomes more effective with larger amount of data such as when multiple businesses share their data to carry out the clustering together. • However, the data may contain sensitive information. Secure K-Means Algorithms • Trusted Third Party (TTP) based solutions – Dwork et al. ( Crypto 2004) Very Efficient No TTP in Real World, Possible security compromise • Data Perturbation techniques – Stanley et al. (BSD 03), Kargupta et al. (ICDM 03) Negligible communication overhead Partial security, Non-invertible transformations used IIIT Hyderabad • Those employing Multiparty Computations – Vaidya et al. (KDD 03), Jha et al. (ESORICS 05) Wright et al. (KDD 05), Inan et al (DKE 07) Complete privacy Highly in-efficient Our Distributed Solution IIIT Hyderabad • We simulate TTP on a set of un-trusted servers over an insecure network. • Secret Sharing is a method of distributing a secret among a group of servers. Proposed Protocol • Protocol consists of two phases – Phase One: Secure Data Distribution – Phase Two: Secure K-Means • Phase One: Secure Storage of data at servers – Selection of an optimal RNS. – Shattering of the user’s private data. Privacy: Server stores only the shattered shares of data. IIIT Hyderabad • Phase Two: Secure K-Means – Initialization – Lloyd Step – Knowledge Revelation Phase Two: Secure K-Means • Clusters are initialized using the shattered shares • Lloyd Step involves iteratively computing the closest centers in a Euclidean space – Secure protocols for division and comparison • Securely evaluate the termination criteria – Send the shattered cluster centers to users who uses the Merge function on it • Privacy: No information is leaked to the servers IIIT Hyderabad – Data for operations such as division secured using randomization – Randomization done so as to secure against possible GCD and factorization based attacks IIIT Hyderabad Overview of the Protocol User 1 User 2 Analysis • Overheads calculated over the naïve TTP based protocol. • Division and Comparison operations introduce communication overhead. – Limited to one round per operation • Traditional approaches uses SMC for this. – Based on OT, a communicational intensive protocol. – O(n2) communication overhead to multiply two vectors (length n) • Limited data expansion IIIT Hyderabad – Eg: 32bit data shattered into 5 shares requires 54bits while traditional SS requires 160bits. Algorithm Properties • We have proposed a highly secure framework using paradigm of secret sharing. • Negligible overheads in simulating algebraic operations. • Achieve efficiency by exploiting the data properties. IIIT Hyderabad • Solution does not demand any trust and the clustering is carried out directly on the encrypted data. Conclusion Broad Objective • Development of secure computational algorithms in computer vision and related areas. – To develop “highly-secure” solutions – To develop “computationally efficient” solutions – To develop solutions to problems with immediate impact • The traditional methods of ensuring privacy are communication and computation expensive. • We show that domain specific knowledge can be incorporated to ensure efficiency while retaining privacy. IIIT Hyderabad • Moreover, our methods do not trade off accuracy. Related Publications Maneesh Upmanyu, Anoop M. Namboodiri, K. Srinathan and C.V. Jawahar; “Blind Authentication - A Secure Crypto-Biometric Verification Protocol” In IEEE-Transactions on Information Forensics and Security (IEEE-TIFS, June 2010) “Efficient Biometric Verification in Encrypted Domain” In Proceedings of 3rd International Conference on Biometrics (ICB 2009) IIIT Hyderabad “Efficient Privacy Preserving Video Surveillance” Proceedings of the 12th International Conference on Computer Vision (ICCV 2009) “Efficient Privacy Preserving K-Means Clustering” Proceedings of the Pacific Asia Workshop on Intelligence and Security Informatics (PAISI 2010) IIIT Hyderabad Thank you for your attention RNS & CRT • Residue Number System (RNS) is an integer using a set of smaller integers. – RNS is defined by a set of k integer constants. {m1, m2, m3, …, mk} – Secret A is represented by k smaller integers. {a1, a2, a3, …, ak} where ai = A modulo mi – This representation is valid as long as 0 < A < M, where M is LCM of mi’s • Chinese Remainder Theorem (CRT) is the method of recovering the integer value from a given set of smaller integers. IIIT Hyderabad – Define Mi = M/mi – Compute ci = Mi x (Mi-1 mod mi) – The above equation is always valid in our system, therefore unique solution exists Shatter & Merge Functions • Shatter function of the private data. : Compute and store the secret shares – Where xi is the ith secret share, and η is a uniform randomness • Merge function : Reconstruct the secret. for different primes Pi’s, secret is IIIT Hyderabad – Given recovered using CRT