Recommender System: Threat Analytics & Secu re Multi-party ...static.tongtianta.site/paper_pdf/3e110fa2-0140-11e9-9a16-00163e08bb86.pdfrecommender system may be subjected to various

See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/320271886

Recommender System: Threat Analytics & Secure Multi-party Computation

Technical Report · April 2016

DOI: 10.13140/RG.2.2.11704.44807

CITATIONS

0READS

68

1 author:

Some of the authors of this publication are also working on these related projects:

Management Information Systems View project

Strategic Management View project

Sumit Chakraborty

Fellow of Indian Institute of Management Calcutta

27 PUBLICATIONS   101 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Sumit Chakraborty on 08 October 2017.

The user has requested enhancement of the downloaded file.

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 1

Recommender System: Threat Analytics & Secure Multi-party Computation

Sumit Chakraborty Fellow, Management Information Systems (Indian Institute of Management Calcutta),

Bachelor of Electrical Engineering (Jadavpur University), India E-mail: [email protected], [email protected]; Phone: 91-9940433441

Abstract : This work deals with the problem of rank computation by a corrupted recommender system. It presents Fair Recommendation Algorithm (FRA) and related complexity analysis and explores cryptographic challenges. FRA assesses and mitigates various types of attacks on a recommender system with the support of an intelligent threat analytics. Secure multi-party computation (SMC) may be an interesting solution for the aforesaid problem from the perspectives of secret sharing, privacy, fairness, correctness, rationality, trust, commitment, integrity, consistency, transparency, accountability, robustness and stability. It is also important to verify authentication, authorization, correct identification, privacy and audit of rank computation by an efficient recommender system. Another critical issue is how to share a secret through threshold cryptographic schema.This work analyzes two test cases with the support of fair recommendation algorithm: (a) ranking in assessment and accreditation of education institutes and (b) rank computation in joint entrance examination (e.g. medical, engineering). This study can be extended to various application domains such as financial service, digital advertising, healthcare, education and corporate governance. Keywords: Recommender System, Shilling attack, Sybil attack, False data injection attack, Knowledge attack, Integrity attack, Basic attack, Rank computation, Secret Sharing, Privacy, Fairness, Correctness, Rationality, Secure Multi-party Computation, Threshold cryptography.

1. INTRODUCTION

Traditionally, a Recommender System is an information system giving suggestionsfor specific set of items in electronic commerce and mobile commerce applications [1]. The suggestions are used in purchasing decision making processes such as what items to buy, what books or online news to read, what songs or music to listen or what movies to watch. An item is an object what the information system recommends to the users.A recommender system is designed with graphical user interface, specific items and core recommendation algorithms to identify or predict a set of useful items for the users or customers or clients or service consumers [2]. The system tries to predict the utility or compare the utility ofsome items and decide what items to recommend. The rating of the raters is used to model a utility function f(u,i) where u is utility for item i. A recommender system computes rank of a set of items based on utility function. The concept can be extended to miscellaneous application domains such as education, financial service, healthcare and corporate governance. A recommender system may be subjected to various types of malicious attacks such as shilling, sybil, false data injection and knowledge attacks [4,11,12,13,14,15]. Many of these works have suggested a set of statistical metrics on ratings to distinguish attack identities from regular identities, attack strategies, impact analysis, online learning algorithms, collaborative filtering algorithms and also attack resistant algorithms. Complex decision making process often requires a scalable, proactive, distributed, intelligent and privacy preserving recommender system. The contributions of this work are as follows. Section 1 starts with introduction which defines the problem of rank computation by a recommender system. It reviews existing literature and analyzes the gaps, states research methodology and contributions of the work. Section 2 presents Fair

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 2

Recommendation Algorithm (FRA) [3,5,6,7,16]. Section 3 calls an intelligent threat analytics and a set of attack resistant algorithms. Section 4 analyzes FRA in terms of security intelligence and computational complexity. Section 5analyzes two test cases with the support of FRA: (a) ranking in assessment and accreditation of education institutes [8,9] and (b) rank computation in joint entrance examination. Section 6 concludes the work by exploring new directions of research on this problem.

2. FAIR RECOMMENDATION ALGORITHM (FRA)

Assumptions: (a) The recommendation algorithm must satisfy the basic requirements of security and privacy from the perspectives of collective intelligence.FRA is basically an algorithmic mechanism. (b) The analytics must explore the risk of all possible threats on a recommender system. (c) Another critical issue is low computation and communication overhead for security intelligence. (d) The recommender system must support scalability and reliability. Agents: Recommender system administrator (A), a set of raters, a set of candidates or entities (E

i,i=1,..,n);

System : Recommender system (R); Objectives: Fair and correct rank computation of the entities; Constraints: Trust, motivation and commitment; Input: Multiple criteria (C

j,j=1,…k);

Strategic moves: Multi-criteria Decision Making (MCDM); Call intelligent threat analytics to assess risk of various malicious attacks on R; Verify security intelligence of R based on the properties of Secure multiparty

computation (SMC); Evaluate the reputation of the raters and system administrator;

Protocol: Authenticate a set of objects or alternatives of decision (E

i,i=1,..,n) through correct

identification; Define a consistent family of criteria (C

j,j=1,..k);

Develop a global preference model; Select appropriate decision support system; Define accountability of a set of authorized raters and communicate them rank

computation mechanism with transparency; Compute score of the entitiesS

i,i=1,..,n= ∑k

j=1w

j.m

jwhere w: weight, m : rating of criterion;

Compute rank of the entities sort Si,i=1,..,n

; Audit fairness and correctness of rank computation;

Rationality of rank computation policy; Optimal set of criteria; Weight assignment; Score / rating evaluation; Sorting algorithm;

Verify risk of malicious attacks on R: shilling attack

push attack; nuke attack;

Evaluate reputation, trust and bias of the raters;[ Ref. algorithm 1] Sybil attack; [ Ref. algorithm 2] false data injection attack

cross validation from authenticated data sources;[ Ref. algorithm 3] basic attack

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 3

random attack; average attack;

Revaluation and recheck;[ Ref. algorithm 4] high / low knowledge attack

bandwagon; reverse bandwagon; segment; love / hate attacks;

Evaluate trust, motivation and commitment;[ Ref. algorithm 5] integrity attack;[ Ref. algorithm 6]

Revelation principle: The system administrator preserves privacy of critical data; Secret sharing through threshold cryptographic schema;

Payment function: Audit business intelligence in terms of incentives received by corrupted agents and

adversaries; The honest agents compute penalty function and charge the corrupted agents;

Output : Rank of candidates or entities;

3. THREAT ANALYTICS This section assesses and mitigates various types of attacks on the recommender system with the support of an intelligent threat analytics. An attack is a concerted effort to bias the outcome of a recommender system by insertion of a large number of profiles using false identities or attack profile. The best attack yields the biggest impact for the least amount of effort. The efforts are required for crafting profiles and gaining knowledge. A high-knowledge attack requires detailed knowledge of the rating distribution; a low-knowledge attack requires system independent knowledge. A robust adaptive and stable recommender system is expected to be protected from following various types of common attacks through algorithms 1-6. 3.1 Algorithm 1 Threat: Shilling attack;

Push attack : promote target item; Nuke attack : demote target item;

Risk assessment: evaluate the quality of recommendation; Detect shilling attacks based on a set of metrics to mine rating patters of the raters

Number of Prediction-Differences Standard Deviation in User’s Ratings Degree of Agreement with Other Users Degree of Similarity with Top Neighbors

Risk mitigation: call influence limiter algorithm which computes reputation of the raters based on scoring rule and loss function. In case of shilling attack, an attacker tries to draw attention to the target items that don’t deserve that attention by influencing a recommender system. For example, the objective of the adversary may be to generate positive recommendations for her own products and poor recommendations for her competitor’s products through shilling attack. An influence-limiting algorithm is expected to protect a recommender system from shilling attack. According to this risk mitigation initiative,

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 4

honest reporting is the dominant strategy for the raters who wish to maximize their influence. The system gives importance to the feedback received from honest and informative raters and reward them based on their performance.

Figure 1: Recommender System’s Mechanism

3.2 Algorithm 2 Threat: Sybil attack; Risk assessment: Detect sybil identities and intrusion of malicious agents associated with the recommender system; Risk mitigation:

trusted explicit and implicit certification; robust authentication protocol; resource testing; incentive based sybil detection game (e.g. auction, discriminatory reward negotiation)

A recommender system is defined by a set of entities, a communication cloud and a set of pipes connecting the entities to the communication cloud. The entities can be partitioned into two subsets: correct and faulty. Each correct entity presents one legitimate identity to other entities of the distributed system. Each faulty entity presents one legitimate identity and one or more counterfeit identities to the other entities. Each identity is an informational abstract representation of an entity that persists across multiple communication events. The entities communicate with each other through messages. A malicious agent may control multiple pseudonymous identities and can manipulate, disrupt or corrupt a recommender system that relies on redundancy. This is known as sybil attack. Sybil attacks may affect fair resource allocation, routing mechanisms, voting, aggregation and storage of distributed data by injecting false data or suppressing critical data. There are various approaches of sybil detection: trusted explicit and implicit certification, robust authentication protocol, resource testing, auction and incentive based sybil detection game [11]. In case of trusted certification, a centralized authority assigns a unique identity to each entity. The centralized authority can verify computing, storage and bandwidth capability of the entities on periodic basis. A local identity (l) accepts the identity (i) of an entity (e) if e presents i successfully to l. An entity may validate the identity of another identity through a trusted agency or other entities or by itself directly. In the absence of a trusted authority, an entity may directly validate the identities of other entities or it may accept identities vouched by other accepted entities. The system

Secure

Multi-party

Computation

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 5

must ensure that distinct identities refer to distinct entities. An entity can validate the identity of other entities directly through the verification of communication, storage and computation capabilities. In case of indirect identity validation, an entity may validate a set of identities which have been verified by a sufficient count of other identities that it has already accepted. But, a group of faulty entities can vouch for Sybil identities.

Figure 2 : Sybil attack on a recommender system

Sybil attack is basically a profile injection attack where a set of profiles are added to the recommender system by an adversary. A profile is a vector of all items, with a rating for each item and null value for unrated items. The basic objective of the attacker is to promote or demote a target item. Figure 4 illustrates Sybil attack on a recommender system by showing authentic user profiles (a to g) and a number of attack profiles (i to m). User h is seeking a prediction for item 7 which is the subject of a nuke attack. The authentic profile recommends positive rating but the attack profiles recommend negative rating of the target item 7. 3.3 Algorithm 3 Threat: False data injection attack; Risk assessment: verify correctness of data input into the recommender system and accountability of the corrupted agents; Risk mitigation: cross validation from authenticated data sources; 3.4Algorithm 4 Threat: Basic attack

random attack : assign rating randomly with minimal knowledge. average attack : assign mean rating for each item based on perception and high degree

of system specific knowledge; cloning attack

Risk assessment: audit bias in distribution of rating data; Risk mitigation:

Collect feedback from the candidates; revaluate and recheck the output of recommender system.

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 6

3.5 Algorithm 5 Threat: Knowledge attack

High knowledge attack: informed attack models requiring high degree of system specific knowledge (e.g. similarity algorithm);

o Popular attack where similarities between users are calculated using Pearson correlation;

Low knowledge attack; Bandwagon attack : associate the attacked item with a small number of frequently

rated items (extension of random attack); reverse bandwagon: assign low rating to the target item associated with widely

disliked items; segment attack : push an item to a targeted group of users with known or easily

predicted preferences; love / hate attacks: very low knowledge simple attack assigning max. / min. rating

to the target item; Risk assessment: audit bias in distribution of rating data; Risk mitigation:

explore motivation, commitment and knowledge of the raters; 3.6 Algorithm 6 Threat: integrity attack; Risk assessment: audit the matching between input data and the data registered into the recommender system; Risk mitigation:

Withdraw input; lodge complain against corruption at top level of system administration;

4. COMPLEXITY ANALYSIS Theorem 1 : The stability and robustness of FRA depends on efficiency of verification mechanism of security intelligence of a recommender system. The security intelligence of FRA is defined with a novel concept of collective intelligence and in terms of a set of properties of secure multi-party computation: authentication, authorization, correct identification, privacy and audit; fairness, correctness, transparency, accountability, trust, non-repudiation, data integrity, reliability and consistency. FRA adopts the concept of influence limiter, trust and reputation system. It is rational to weight the contribution of each user towards a prediction through a measure of reputation. The reputation value is boosted when a rating is correctly estimated and is reduced which it fails to do so. The influence limiter is based on trust and reputation to make intelligent recommendations. The reliability of a profile to deliver accurate recommendations in the past should be taken into account by FRA. FRA must address correct identification, authentication, authorization, privacy and audit for appropriate access control of each stakeholder associated with the recommender system. For any secure service, the system should ask the identity and authentication of one or more agents involved in a communication. The agents of the same trust zone may skip authentication but it is essential for all sensitive communication across different trust boundaries. After the identification and authentication, FRA should address the issue of authorization. The system should be configured in

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 7

such a way that an unauthorized agent cannot perform any task out of scope. The system should ask the credentials of the requester; validate the credentials and authorize the agents to perform a specific task as per agreed protocol. Each agent should be assigned an explicit set of access rights according to role. Privacy is another important issue; an agent can view only the information according to authorized access rights. A protocol preserves privacy if no agent learns anything more than its output; the only information that should be disclosed about other agent’s inputs is what can be derived from the output itself. The privacy of data may be preserved in different ways such as adding random noise to data, splitting a message into multiple parts randomly and sending each part to an agent through a number of parties hiding the identity of the source, controlling the sequence of passing selected messages from an agent to others through serial or parallel mode of communication, dynamically modifying the sequence of events and agents through random selection and permuting the sequence of messages randomly. The agents must commit the confidentiality of strategic data used in the recommender system. The system administrator must be able to audit the efficiency of ranking mechanism at anytime in terms of fairness and correctness. There are some other important parameters of security intelligence: fairness, correctness, transparency, accountability and trust. FRA is expected to verify correctness of rank computation. The fairness of FRA is associated with the rationality of rank computation policy, optimal set of criteria, rational weight assignment and rating evaluation. Fairness ensures that something will or will not occur infinitely often under certain conditions. The mechanism must ensure the accountability and responsibility of the agents in access control, data integrity and non-repudiation. The transparency of FRA is associated with communication protocols, revelation principle and verification mechanisms. In fact, the issues of correctness, fairness, transparency and accountability are all interlinked. FRA calls threat analytics: assesses risks of single or multiple threats on the recommender system such as false data injection, Sybil, shilling: push and pull, nuke attack, corruption in secret sharing and information leakage, basic attack : random / average attack, high / low knowledge attack : bandwagon, reverse bandwagon, segment and love / hate attacks. The computational complexity of FRA mainly depends on the cost of rank computation and threat analytics. Theorem 2 : FRA ensures privacy through a well-defined revelation principle, SMC and signcryption. Privacy is the primary concern of FRA; the issue can be addressed utilizing the concept of cryptography including secure multiparty computation. The fundamental objectives of cryptography are to provide confidentiality, data integrity, authentication and non-repudiation. Cryptography ensures privacy and secrecy of information through encryption methods. The sender (S) encrypts a message (m) with encryption key and sends the cipher text (c) to the receiver (R’). R’ turns c back into m by decryption using secret decryption key. In this case, an adversary may get c but cannot derive any information. R’ should be able to check whether m is modified during transmission. R’ should be able to verify the origin of m. S should not be able to deny the communication of m. The raters may send rating in the form of signcrypted messages to the recommender system. Traditional signature-then-encryption is a two step approach. At the sending end, the sender signs the message using a digital signature and then encrypts the message. The receiver decrypts the cipher text and verifies the signature. The cost for delivering a message is the sum of the cost of digital signature and the cost of encryption. Signcryption is a public key primitive that fulfills the functions of digital signature and public key encryption in a logically single step and the cost of delivering a signcrypted message is significantly less than the cost of signature-then-encryption approach. FRA may use two types of key based algorithms - symmetric and public key. Symmetric key encryption scheme provides secure communication for a pair of communication partners; the sender

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 8

and the receiver agree on a key k which should be kept secret. In most cases, the encryption and decryption key are same. In case of asymmetric or public-key algorithms, the key used for encryption (public key) is different from the key used for decryption (private key). The decryption key cannot be calculated from the encryption key at least in any reasonable amount of time. The widely-used public–key cryptosystem are RSA cryptosystem (1978), Elgamal’s cryptosystem (1985) and Paillier’s cryptosystem (1999). FRA may preserve privacy of critical strategic data through SMC. Two or more agents want to conduct a computation based on their private inputs but neither of them wants to share its proprietary data set to other. The objective of secure multiparty computation (SMC) is to compute with each party’s private input such that in the end only the output is known and the private inputs are not disclosed except those which can be logically or mathematically derived from the output. In case of secure multi-party computation, a single building block may not be sufficient to do a task; a series of steps should be executed to solve the given problem. Such a well-defined series of steps is called a SMC protocol. A SMC protocol is expected to satisfy a set of properties – privacy, correctness, independence of inputs, guaranteed output delivery and fairness. A protocol ensures correctness if each party receives correct output. Corrupted (or malicious) parties select their inputs independently of the inputs of honest parties and honest parties must receive their output. Corrupted parties should receive their outputs if and only if the honest parties receive their outputs and this ensures fairness of the protocol. FRA may consider two different models for secure multi-party computation – semi-honest model and malicious model. A semi-honest party follows the protocol properly with correct input. But after the execution of the protocol, it is free to use all its intermediate computations to compromise privacy. A malicious party does not need to follow the protocol properly with correct input; it can enter the protocol with an incorrect input. A third party may exist in a protocol. A trusted third party is given all data; it performs the computation and delivers the result. In some SMC protocols, an untrusted third party is used to improve efficiency. Threshold cryptography may be also useful to preserve privacy (refer test case 2).

5. RESEARCH METHODOLOGY : EXPERIMENTAL RESULTS

This work deals with the problem of rank computation by a corrupted recommender system. The research methodology is primarily focused on logical and analytical case based reasoning of two test cases : (a) ranking and accreditation of education institutes and also digital advertising and (b) rank computation in joint entrance examination (e.g. medical, engineering). 5.1 TEST CASE 1 - Ranking and Accreditation of Education Institutes Fair Recommendation Algorithm I (FRA-I) Agents: Recommender system administrator (A), a set of raters, a set of entities i.e. education institutes (E

i,i=1,..,n);

System : Assessment and accreditation system (R); Objectives: Fair and correct rank computation of the education institutes; Constraints: Trust, motivation and commitment; Input: Multiple criteria (C

j,j=1,…k);

Strategic moves: Multi-criteria Decision Making (MCDM); Call intelligent threat analytics to assess risk of various malicious attacks on R; Verify security intelligence of R based on the properties of secure multiparty

computation (SMC); Evaluate the reputation of the raters and system administrator;

Protocol:

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 9

Authenticate a set of objectsi.e. education institutes (Ei,i=1,..,n

) through correct identification;

Define a consistent family of criteria (Cj,j=1,..k

); Develop a global preference model; Select appropriate decision support system; Define accountability of a set of authorized raters and communicate them rank

computation mechanism with transparency; Compute score of the entities S

i,i=1,..,n= ∑k

j=1w

j.m

j where w: weight, m : rating of criterion;

Compute rank of the education institutes sort Si,i=1,..,n

; Audit fairness and correctness of rank computation rationally; Verify risk of malicious attacks on R:

shilling attack : push and nuke attack; Evaluate reputation and bias of the raters periodically through cross

validation. False profile data injection attack

Verify authenticity of profile data of each education institute. Get feedback from various stakeholders such as students, parents and

faculties apart from the inspectors. Low knowledge attack

Audit casual commitment and motivation in proper knowledge management, research, innovation and creativity at each education institute;

Evaluate real contributions in academics, industry and commerce and growth of the society;

Identify artificial mechanical approach of various education institutes to secure high rank ignoring the basics of conceptual deep learning;

Revelation principle: The system administrator preserves privacy of critical strategic data of rating and

profile of each educational institute; The system administrator reveals assessment and accreditation policy and mechanism

publicly with transparency; Payment function:

Audit business intelligence in terms of incentives received by corrupted agents and adversaries;

The honest agents compute penalty function and charge the corrupted agents; Output : Ranking and accreditation of educational institutes. In the first test case, malicious broadcast is a real threat to the digital advertising world, education and financial service sector. Existing manuals of schools, colleges and universities should address the risk of various malicious threats on the assessment and accreditation system [8,9]. Today’s broadcast is closely associated with advertising as a recommender system. But, there is risk of shilling attack in the form of push and nuke attacks where the rating of target items are increased and lowered successively. An attacker can draw attention to items that do not deserve the attention of consumers by manipulating rank computation by a recommender system. The advertising world may be digitally divided with a flavor of revenge and retaliation due to zero or low investment on advertising by the corporate world. A corrupted broadcasting system may be involved in brand dilution of a good company through baseless, mischievous and false propaganda. Alternatively, the broadcasting system can push a set of targeted items of poor quality and brand to the public through fraudulent adwords, euphemism and attractive presentation of the popular brand ambassadors. But after the disclosure of the information on such types of malicious attacks, the recipients may lose

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 10

their trust in the adwords of the digital world in future. If the recipients sense flaws in digital advertising, the system administrator must verify the correctness, fairness and transparency of the system through analytics on ad slot allocation, content of adwords, exposure time and frequency, customization, delivery, click rate, and impression. An efficient influence limiting algorithm can promote a manipulation resistant recommender system where honest reporting is the optimal strategy for the raters to maximize their influence in rank computation. 5.2 TEST CASE 2 – Corrupted Rank Computation in Joint Entrance Examination Fair Recommendation Algorithm II (FRA-II) Agents: Recommender system administrator (A), a set of raters, a set of candidates i.e. students (E

i,i=1,..,n);

System :Joint Entrance Examination ranking system (R); /* R is assumed to be a specific type of recommender system */ Objectives: Fair and correct rank computation of the candidates; Constraints: Trust, honesty, motivation and commitment; Input: Multiple criteria (C

j,j=1,…k);

Strategic moves: Multi-criteria Decision Making (MCDM); Call intelligent threat analytics to assess risk of various malicious attacks on R; Verify security intelligence of R based on the properties of secure multiparty

computation (SMC); Evaluate the reputation of the raters and system administrator;

Protocol: Authenticate a set of objects i.e. candidates(E

i,i=1,..,n) through correct identification;

Define a consistent family of criteria (Cj,j=1,..k

) i.e. intelligent questions to evaluate the basic concept and aptitude of the candidates;

Develop a global preference model; Select appropriate decision support system; Define accountability of a set of authorized raters and communicate them rank

computation mechanism with transparency; Compute score of the entities S

i,i=1,..,n= ∑k

j=1w

j.m

j where w: weight, m : rating of criterion;

Compute rank of the candidates sort Si,i=1,..,n

; Audit fairness and correctness of rank computation; Verify risk of malicious attacks on R:

shilling attack : push and nuke attack; Evaluate reputation and bias of the raters periodically through cross

validation; False profile data injection attack

Verify authenticity of profile data of each candidate; Get feedback from alternative sources and stakeholders;

Integrity attack in online test Restrict modification of answers or online registration data of some

candidates by adversaries in exchange of incentives through discriminatory treatment;

Low knowledge attack Evaluate the intelligence and correctness of test papers; Identify artificial mechanical approach of some candidates to secure high

rank ignoring the basics of conceptual deep learning;

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 11

Audit casual commitment and motivation in proper knowledge management, innovation and creativity;

Revelation principle: The system administrator preserves privacy of critical strategic data of question

papers or test papers before the start of the joint entrance examination; Form an independent neutral expert panel for setting test papers; Secret sharing through (k,n) threshold cryptographic schema;

Data D is divided into n pieces D1,…, D

n

knowledge of any k or more Di pieces makes D easily

reconstructable; even complete knowledge of (k-1) pieces reveals absolute no

information of D; Restrict the disclosure of test papers through dishonest channels;

The system administrator reveals assessment and evaluation policy and mechanism publicly with transparency.

Payment function: Audit business intelligence in terms of incentives received by corrupted agents and

adversaries; The honest agents compute penalty function and charge the corrupted agents;

Output:Rank of candidates FRA-II algorithm clearly shows the importance of correctness and fairness of rank computation for technology management and high quality healthcare services for the sustainability of human civilization in coming future. In the second case, the rank computation by a recommender system for joint entrance examination (e.g. medical science or engineering) requires a scientific approach to streamline the process : (a) As-is system and process analysis of joint entrance examination, (b) identification of gaps of the existing system and (c) define to-be system and process. For instance, the design of the questions for joint entrance medical examination is expected to be innovative thought provoking, intelligent and interesting application oriented from the perspectives of medical science, healthcare and life-science. The basic objective is expected to test the concept and aptitude of the candidates, not only their memorizing skill. The design of questions is expected not to be mechanical, vague, boring or just a quiz contest. The paper should be set by an independent neutral expert panel. It is easy to check Multiple Choice Questions (MCQ) through computers; but it is not adequate to evaluate the merit, intelligence, understanding, thinking, reasoning and decision making capabilities of to-be doctors. It is really essential to think that the successful candidates passing the joint entrance medical examination will have to take care of precious human life in future. Another important issue in the second test case is secret sharing; it is essential to explore new cryptographic challenges such as threshold cryptography. It is a debatable issue whether (k,n) threshold cryptographic schema is really useful to preserve the privacy of test papers in this test case.Threshold cryptographic schemes are generally applicable to a group of mutually suspicious individuals with conflicting interest who must cooperate [9]. Let us consider a specific case of complex, hard, difficult and tricky joint entrance medical examination question paper. Apparently, it seems an attempt to improve the quality of medical education. But, the privacy of question paper is a very important factor in this context even in case of online test. The questions are expected not to be disclosed to a set of candidates through various channels such as coaching institutes or private tutors. Otherwise, the system will lose fairness, competitiveness, rationality and correctness. Only, the candidates from rich class may be able to study medical and engineering education in the coming future. There should be no conflict between business intelligence, fairness, correctness and transparency of the system i.e. ‘bajro atuni faska gero’. An efficient recommender system is expected not to be crippled with any economic pressure. It is also essential to revise and redesign the text

Reference of document : Technical Report TR/RCCRS V2.0 DATED 01.04.2016 Page 12

books of Physics, Chemistry and Biology, particularly various mechanisms of human biological system should be discussed with more clarity, transparency and more depth and breadth. This study can be extended to the rank computation in joint entrance engineering examination also.

6. CONCLUSION

Malicious attacks on ranking and rating computation of a recommender system represent a vibrant and changing research domain. This work has found a number of challenges and future research directions: intelligent threat analytics, rational trust modeling, new algorithms and heuristics for risk assessment and risk mitigation, computational and communication complexity, secure multiparty computation, secret sharing through threshold cryptographic schema; scalable, proactive, distributed and privacy preserving recommender system. It provides a systematic view of rank computation of an efficient and intelligent recommender system and a number of promising future research directions. It is just a modest attempt towards exploring the problem of rank computation by recommender systems and much more research are needed to unlock the full potential of fair rank computation in various application domains such as financial service (stock and bond market), healthcare (hospitals), education (journal, institutes), HR (performance measurement) and corporate governance (company rating, credit rating).

REFERENCES 1. Konstan, J.A. 2004. Introduction to Recommender Systems: Algorithms and Evaluation. ACM Transactions on Information Systems (TOIS), 22(1):1–4. 2. Resnick, P., Varian, H.R. 1997. Recommender Systems.Communications of the ACM, 40(3):56–58. 3. Roy, B., Bouyssou, D. 1993. Aide multicritére á la décision: méhodesetcas. Economica, Paris. 4. Price, B, Messinger, P.R. 2005. Optimal Recommendation Sets: Covering Uncertainty over User Preferences. Proc. of the National Conference on Artificial Intelligence. 5. Van Roy, B., Yan, X.. 2009. Manipulation-resistant collaborative filtering systems. In: RecSys 09:Proceedings of the third ACM conference on Recommender systems, pp. 165–172. ACM,New York, USA. 6. Goldreich, O.1998. Secure multi-party computation. 7. Lindell, Y. 2003. Composition of secure multi-party protocols a comprehensive study. Springer. 8. NAAC. 2017. Manual for colleges. Bengaluru, July. 9. NAAC. 2017. Manual for universities. Bengaluru, July. 10. Shamir, A.1979. How to share a secret.ACM. 11. Pal,A.K., Nath, D. and Chakraborty,S. 2010. A Discriminatory Rewarding Mechanism for Sybil Detection with Applications to Tor, WASET. 12. Chirita, P., Nejdl, W and Zamfir, C. 2005. Preventing shilling attacks in online recommender systems.WIDM’05, Bremen, Germany. 13. Resnick, P. and Sami, R. 2007. The Influence Limiter: Provably Manipulation Resistant Recommender Systems. Recsys’2007, Minnesota, USA. 14. Burke, R., Mobasher, B. and Williams, C. 2006. Classification features for attack detection in collaborative recommender systems. In: Proceedings of the 12th International Conference on Knowledge Discovery and Data Mining, pp. 17–20. 15. Burke, R., Mobasher, B., Zabicki, R. and Bhaumik, R. 2005. Identifying attack models for secure recommendation. In: Beyond Personalization: A Workshop on the Next Generation of Recommender Systems. 16. Adomavicius, G., Kwon, Y. 2007. New Recommendation Techniques for Multi-Criteria Rating Systems. IEEE Intelligent Systems, 22(3):48–55.

View publication statsView publication stats