Recent research has identified significant vulnerabilities in recommender systems. Shilling attacks, in which attackers introduce biased ratings in order to influence future recommendations, have been shown to be effective against collaborative filtering algorithms. We postulate that the distribution of item ratings in time can reveal the presence of a wide range of shilling attacks given reasonable assumptions about their duration. To construct a time series of ratings for an item, we use a window size of k to group consecutive ratings for the item into disjoint windows and compute the sample average and sample entropy in each window. We derive a theoretically optimal window size to best detect an attack event if the number of attack profiles is known. For practical applications where this number is unknown, we propose a heuristic algorithm that adaptively changes the window size. Our experimental results demonstrate that monitoring rating distributions in time series is an effective approach for detecting shilling attacks.

Collaborative filtering techniques have become popular in the past several years as an effective way to help people deal with information overload. An important security concern in traditional recommendation systems is that users disclose information that may compromise their individual privacy when providing ratings. Randomized perturbation schemes have been proposed to disguise user ratings while still producing accurate recommendations. However, recent research has suggested that perturbation schemes might not be able to preserve privacy as much as has been believed. We propose two data reconstruction methods that derive original private information from disguised data in existing perturbation collaborative filtering schemes. One method is based on k-means clustering and the other uses singular value decomposition (SVD). We have conducted theoretical and experimental analysis on the difference between original data and reconstructed data. Our experiments show that both methods can derive a considerable amount of original information. This study helps to determine an empirical trade-off between recommendation accuracy and user privacy in perturbation schemes.

We use a low-dimensional linear model to describe the user rating matrix in a recommendation system. A non-negativity constraint is enforced in the linear model to ensure that each user's rating profile can be represented as an additive linear combination of canonical coordinates. In order to learn such a constrained linear model from an incomplete rating matrix, we introduce two variations on Non-negative Matrix Factorization (NMF): one based on the Expectation-Maximization (EM) procedure and the other a Weighted Nonnegative Matrix Factorization (WNMF). Based on our experiments, the EM procedure converges well empirically and is less susceptible to the initial starting conditions than WNMF, but the latter is much more computationally efficient. Taking into account the advantages of both algorithms, a hybrid approach is presented and shown to be effective in real data sets. Overall, the NMF-based algorithms obtain the best prediction performance compared with other popular collaborative filtering algorithms in our experiments; the resulting linear models also contain useful patterns and features corresponding to user communities.

An important security concern with traditional recommendation systems is that users disclose information that may compromise their individual privacy when providing ratings. A randomization approach has been proposed to disguise user ratings while still producing accurate recommendations. However, recent research has suggested that a significant amount of original private information can be derived from perturbed data in a randomization scheme. We suggest that a main limitation of the existing randomization approach is that perturbation is item-invariant--each item has a same perturbation variance. Based on this observation, we introduce a two-way communication privacypreserving scheme in which users perturb their ratings for each item based on the server's guidance instead of using an item-invariant perturbation. Compared to the existing randomization approach, our new scheme can help users disclose much less private information at the same recommendation accuracy level.

Collaborative filtering techniques have become popular in the past decade as an effective way to help people deal with information overload. Recent research has identified significant vulnerabilities in collaborative filtering techniques. Shilling attacks, in which attackers introduce biased ratings to influence recommendation systems, have been shown to be effective against memory-based collaborative filtering algorithms. We examine the effectiveness of two popular shilling attacks (the random attack and the average attack) on a model-based algorithm that uses Singular Value Decomposition (SVD) to learn a low-dimensional linear model. Our results show that the SVD-based algorithm is much more resistant to shilling attacks than memory-based algorithms. Furthermore, we develop an attack detection method directly built on the SVD-based algorithm and show that this method detects random shilling attacks with high detection rates and very low false alarm rates.

This paper describes a hybrid negotiation strategy mechanism using a strategy pool framework that allows negotiation agents more flexibility and robustness in an automated negotiation system. The strategy pool framework supports: a) dynamically assigning an appropriate negotiation strategy to a negotiation agent according to the current negotiation environment and b) creating new negotiation rules by learning from past negotiations. Learning forms we use here for the framework are Feed Forward Back Propagation (FFBP) neural networks and multidimensional inter-transaction association rules mining.