A Robust Bayesian Probabilistic Matrix Factorization Model for Collaborative Filtering Recommender Systems Based on User Anomaly Rating Behavior Detection

( Hongtao Yu ),( Lijun Sun ),( Fuzhi Zhang ) 한국인터넷정보학회 2019 KSII Transactions on Internet and Information Syst Vol.13 No.9

Collaborative filtering recommender systems are vulnerable to shilling attacks in which malicious users may inject biased profiles to promote or demote a particular item being recommended. To tackle this problem, many robust collaborative recommendation methods have been presented. Unfortunately, the robustness of most methods is improved at the expense of prediction accuracy. In this paper, we construct a robust Bayesian probabilistic matrix factorization model for collaborative filtering recommender systems by incorporating the detection of user anomaly rating behaviors. We first detect the anomaly rating behaviors of users by the modified K-means algorithm and target item identification method to generate an indicator matrix of attack users. Then we incorporate the indicator matrix of attack users to construct a robust Bayesian probabilistic matrix factorization model and based on which a robust collaborative recommendation algorithm is devised. The experimental results on the MovieLens and Netflix datasets show that our model can significantly improve the robustness and recommendation accuracy compared with three baseline methods.

Toward Trustworthy Social Network Services: A Robust Design of Recommender Systems

노기섭,오하영,이규행,김종권 한국통신학회 2015 Journal of communications and networks Vol.17 No.2

In recent years, electronic commerce and online social networks (OSNs) have experienced fast growth, and as a result, recommendation systems (RSs) have become extremely common. Accuracy and robustness are important performance indexes that characterize customized information or suggestions provided by RSs. However, nefarious users may be present, and they can distort information within the RSs by creating fake identities (Sybils). Although prior research has attempted to mitigate the negative impact of Sybils, the presence of these fake identities remains an unsolved problem. In this paper, we introduce a new weighted link analysis and influence level for RSs resistant to Sybil attacks. Our approach is validated through simulations of a broad range of attacks, and it is found to outperform other state-of-the-art recommendation methods in terms of both accuracy and robustness.

Robust Recommendation Algorithm based on Metadata Fusion

Gao Feng 보안공학연구지원센터 2014 International Journal of u- and e- Service, Scienc Vol.7 No.6

The metadata information of users and items for enhancing the recommendation system robustness has important valuable. Following this design philosophy, this paper first presents the user suspects assessment strategy based on Probabilistic Latent Semantic Analysis, the user suspected sexual and generic items such as meta-information to model parameters and Logistic Regression way into Bayesian probabilistic matrix factorization (BPMF) model, and then proposes Metadata-enhanced Variational Bayesian Matrix Factorization (MVBMF), designed a model of incremental learning strategy based on robust linear regression, in order to reduce the demand for model rebuilding. Experimental results show that MVBMF can effectively defend against shilling attacks and also has a high level of performance for strong and weak generalization.

Robust Online Filter Recommended Algorithm based on Attack Profile

Gao Feng 보안공학연구지원센터 2014 International Journal of Security and Its Applicat Vol.8 No.4

소셜 트러스트 클러스터 효과를 이용한 견고한 추천 시스템 설계 및 분석

노기섭,오하영,이재훈 한국정보보호학회 2018 정보보호학회논문지 Vol.28 No.1

A Recommender System (RS) is a system that provides optimized information to users in an over-supply situation. Thekey to RS is to accurately predict the behavior of the user. The Matrix Factorization (MF) method was used for thisprediction in the early stage, and according to the recent SNS development, social information is additionally utilized toimprove prediction accuracy. In this paper, we use RS internal trust cluster, which was overlooked in previous studies, tofurther improve performance and analyze the characteristics of trust clusters. 추천시스템(Recommender System, RS)는 정보 과잉 공급 상태에서 사용자들에게 최적화된 정보를 제공하는시스템이다. RS의 핵심은 사용자의 행동 결과를 정확하게 예측하는 것이다. 이러한 예측을 위해 MatrixFactorization (MF) 방식이 초기에 사용되었으며, 최근 SNS의 발달에 따라 Social Information을 추가적으로활용하여 예측 정확도를 높이고 있다. 본 논문에서는 기존 연구에서 간과 되었던 RS 내부 trust cluster를 이용하여 추가적으로 성능을 향상시키고, trust cluster의 특성에 대하여 분석한다. 기존 방법론 3가지와 비교한 결과 본논문에서 제안하는 방식이 가장 높은 정확도를 보임을 확인하였다.

Shilling Attacks Against Memory-Based Privacy-Preserving Recommendation Algorithms

( Ihsan Gunes ),( Alper Bilge ),( Huseyin Polat ) 한국인터넷정보학회 2013 KSII Transactions on Internet and Information Syst Vol.7 No.5

Privacy-preserving collaborative filtering schemes are becoming increasingly popular because they handle the information overload problem without jeopardizing privacy. However, they may be susceptible to shilling or profile injection attacks, similar to traditional recommender systems without privacy measures. Although researchers have proposed various privacy-preserving recommendation frameworks, it has not been shown that such schemes are resistant to profile injection attacks. In this study, we investigate two memory-based privacy-preserving collaborative filtering algorithms and analyze their robustness against several shilling attack strategies. We first design and apply formerly proposed shilling attack techniques to privately collected databases. We analyze their effectiveness in manipulating predicted recommendations by experimenting on real data-based benchmark data sets. We show that it is still possible to manipulate the predictions significantly on databases consisting of masked preferences even though a few of the attack strategies are not effective in a privacy-preserving environment.

시빌 유형을 고려한 견고한 추천시스템

노태완(Taewan Noh),오하영(Hayoung Oh),노기섭(Giseop Noh),김종권(Chongkwon Kim) 한국정보과학회 2015 정보과학회 컴퓨팅의 실제 논문지 Vol.21 No.10

최근 인터넷의 급 성장과 함께 사용자들은 물건이나 영화, 음악 등을 구매 할 때 여러 가지 추천 사이트를 활용한다. 하지만 이러한 추천 사이트에는 악의적으로 아이템의 평점을 높이거나 낮추려는 악의적인 사용자(Sybil)들이 존재할 수 있으며, 추천시스템에 영향을 끼쳐 일반 사용자들에게 부정확한 결과를 추천할 수 있다. 본 논문에서는 사용자들이 생성하는 평점들을 일반적인 평점과 일반적이지 않은 평점으로 구분하고, 상태 정보를 재정립 및 활용하여 악의적 사용자의 영향력을 최소화 하는 추천 알고리즘을 제안한다. 특히, 현재 추천시스템에서의 문제가 되고 있는 3가지 공격모델의 개별 특성을 고려하여 시빌 유형에 견고한 추천 시스템을 처음으로 제안한다. 제안하는 기법의 성능을 입증하기 위해 실제 데이터를 직접 수집(crawling)하여 성능분석결과 제안하는 기법의 성능이 기존 알고리즘과는 다르게 공격 크기 및 종류에 상관 없이 좋은 성능을 보이는 것을 확인 하였다. With a rapid development of internet, many users these days refer to various recommender sites when buying items, movies, music and more. However, there are malicious users (Sybil) who raise or lower item ratings intentionally in these recommender sites. And as a result, a recommender system (RS) may recommend incomplete or inaccurate results to normal users. We suggest a recommender algorithm to separate ratings generated by users into normal ratings and outlier ratings, and to minimize the effects of malicious users. Specifically, our algorithm first ensures a stable RS against three kinds of attack models (Random attack, Average attack, and Bandwagon attack) which are the main recent security issues in RS. To prove the performance of the method of suggestion, we conducted performance analysis on real world data that we crawled. The performance analysis demonstrated that the suggested method performs well regardless of Sybil size and type when compared to existing algorithms.

Shilling Attack Detection Algorithm based on Non-random-missing Mechanism

Man Li 보안공학연구지원센터 2014 International Journal of Security and Its Applicat Vol.8 No.6

Besides unsupervised feature, universality serves as another important factor determining the practical value of attack detection technology. Considering the difficulty of possessing both features for the existing attack detection techniques, this paper reveals the latent factors invoking missing ratings under the non-random-missing mechanism and further combines these latent factors with Dirichlet process in the framework of probabilistic generative model, thus proposes the Latent Factor Analysis for Missing Ratings(LFAMR)model. Based on performing user clustering with this model, this paper achieves the goal of attack detection by presenting the method for identifying attack cluster in ideal situation. Experimental results show that comparing with the existing detection techniques, LFAMR is more universal and unsupervised, and it can effectively detect shilling attacks of typical types and their derivatives even in lack of the apriori inputs such as user cluster numbers.