## Collaborative Filtering with Privacy (2002)

### Cached

### Download Links

- [www.cs.berkeley.edu]
- [jamesthornton.com]
- [www.cs.berkeley.edu]
- [www-kiv.zcu.cz]
- [www-kiv.zcu.cz]
- [www.comp.nus.edu.sg]
- DBLP

### Other Repositories/Bibliography

Citations: | 123 - 9 self |

### BibTeX

@MISC{Canny02collaborativefiltering,

author = {John Canny},

title = {Collaborative Filtering with Privacy},

year = {2002}

}

### Years of Citing Articles

### OpenURL

### Abstract

Server-based collaborative filtering systems have been very successful in e-commerce and in direct recommendation applications. In future, they have many potential applications in ubiquitous computing settings. But today's schemes have problems such as loss of privacy, favoring retail monopolies, and with hampering diffusion of innovations. We propose an alternative model in which users control all of their log data. We describe an algorithm whereby a community of users can compute a public "aggregate" of their data that does not expose individual users' data. The aggregate allows personalized recommendations to be computed by members of the community, or by outsiders. The numerical algorithm is fast, robust and accurate. Our method reduces the collaborative filtering task to an iterative calculation of the aggregate requiring only addition of vectors of user data. Then we use homomorphic encryption to allow sums of encrypted vectors to be computed and decrypted without exposing individual data. We give verification schemes for all parties in the computation. Our system can be implemented with untrusted servers, or with additional infrastructure, as a fully peer-to-peer (P2P) system. 1

### Citations

1013 | Empirical analysis of predictive algorithms for collaborative filtering
- Breese, Heckerman, et al.
- 1998
(Show Context)
Citation Context ... database, the ratings from the linear model are as good as the best current algorithms. In a later section we compare it with neighborhood methods using surveys from Herlocher [7] and Breese et. al. =-=[2]-=-. We construct the k-dimensional linear space A that best approximates the user preference matrix P in a least-squares sense. Assume A is represented as a roworthonormal matrix A ∈ Rk×m . Now k ≤ m wh... |

935 | Oceanstore: An architecture for global-scale persistent storage
- Kubiatowicz, Bindel, et al.
- 2000
(Show Context)
Citation Context ...ations exist in the Groove system (www.groovenetworks.com). It is not clear whether Groove scales to thousands or even hundreds of sites. But other systems under development do. The Oceanstore system =-=[8]-=- is designed to provide global-scale services, potentially to millions of sites. Similar services are under development as part of the JXTA peer-to-peer API within Java. We also assume the existence o... |

669 |
Completeness theorems for non-cryptographic fault-tolerant distributed computation
- Ben-Or, Goldwasser, et al.
- 1988
(Show Context)
Citation Context ...fication is not included in it). We assume that a fraction α > 1/2 of the users’ machines are uncorrupted. Our protocol proceeds in rounds, and we model corruption as a static adversary on each round =-=[1, 12]-=-. That is, the adversary can choose which machines to corrupt at the start of a round, but this choice stays fixed throughout the round. We believe this model is realistic. First of all, in our settin... |

524 |
An algorithmic framework for performing collaborative filtering
- Herlocker, Konstan, et al.
- 1999
(Show Context)
Citation Context ...ts with the “Eachmovie” database, the ratings from the linear model are as good as the best current algorithms. In a later section we compare it with neighborhood methods using surveys from Herlocher =-=[7]-=- and Breese et. al. [2]. We construct the k-dimensional linear space A that best approximates the user preference matrix P in a least-squares sense. Assume A is represented as a roworthonormal matrix ... |

220 | A Secure and Optimally Efficient Multi-Authority Election Scheme
- Cramer, Gennaro, et al.
- 1997
(Show Context)
Citation Context ...hat is practical if both n and m are at the low to middle part of their ranges. Users’ computers perform all the computation in the method. It would be simpler to do this using a set of servers as in =-=[5]-=-, but this would not achieve our goal of a community-based system. Our peer-to-peer version includes additional verification steps that [5] does not (their protocol is “publicly verifiable” but verifi... |

216 | Application of dimensionality reduction in recommender systems–a case study
- Sarwar, Karypis, et al.
- 2000
(Show Context)
Citation Context ...we believe makes economic and practical sense. It is interesting also as plausible application of encrypted multi-party computation. Collaborative filtering using SVD is not new, and was described in =-=[14]-=-. But that paper used simple inner products to generate recommendations, while we use the maximum likelihood formulation of section 2.1, which is novel. This improves the mean-square error of our meth... |

190 |
A threshold cryptosystem without a trusted party
- Pedersen
- 1991
(Show Context)
Citation Context ...e method, with use ideas from the voting algorithm of Cramer, Gennaro, and Schoenmakers [5]. For initial key generation, we assume either an honest dealer, or the distributed key generation scheme of =-=[9]-=-. While we can use Pedersen directly, Cramer et al.’s [5] is a server-based protocol. For our peer-to-peer application, we needed some modifications and extensions to [5]. The main extension is the da... |

171 | Collaborative filtering by personality diagnosis: A hybrid memory-and model-based approach
- Pennock, Horvitz, et al.
- 2000
(Show Context)
Citation Context ...more complicated to analyze other schemes. But schemes which do not create an intermediate model like ours are probably very dangerous. For instance, Pearson correlation [7] and personality diagnosis =-=[10]-=- use the entire user dataset to generate new recommendations. What’s more, Pearson correlation makes use of a subset of “neighbors” of the current user who have rated several of the same items. The ne... |

100 |
Diffusion of Innovations, The. Fourth Edition
- Rogers
- 1995
(Show Context)
Citation Context ...le but important sociological disadvantage. Today’s collaborative filtering algorithms are all based on ratings from the most similar users to a given user. In the language of diffusion of innovation =-=[13]-=-, this is called homophilous diffusion. Homophilous diffusion allows rapid diffusion of innovations within socio-economic groups. But diffusion throughout society requires heterophilous diffusion, whe... |

91 |
Computational Methods in Optimization
- Polak
- 1970
(Show Context)
Citation Context ...ror of our method over [14]. We cannot use a “black-box” SVD algorithm with the limitations of encrypted computation, so we derive an iterative SVD using the conjugate gradient method of PolakRibiere =-=[11]-=-. This gives us an iteration with only vector additions of user data. The derivation is a standard application of numerical techniques, so we include it as Appendix I. For the cryptographic portion of... |

67 | Efficient multiparty computations secure against an adaptive adversary
- Cramer, Damg˚ard, et al.
- 1999
(Show Context)
Citation Context ...the round. Even if we could justify an adaptive adversary, it would not be practical to defend against it. The best multi2party protocols for adaptive adversaries have complexity of O(n3 ) or higher =-=[4]-=-. This is well outside the realm of possibility for a large-scale peer-to-peer community. Fortunately, the static adversary does model the risks to our protocol sensibly. This adversary model allows u... |

8 |
Damg˚ard: Zero Knowledge for Finite Field Arithmetic or: Can Zero Knowledge be for Free?, manuscript
- Cramer, I
- 1997
(Show Context)
Citation Context ...ification section 3.5, which uses sampling and a trusted source of random bits to allow clients to compute reliably with mostly-trustworthy peers. Finally we modified the multiplication protocol from =-=[3]-=- for ZKP of products to work for squares (see Appendix II). The resulting non-interactive proofs are 7 integers rather than 10 integers long. This reduces the overall computation and communication cos... |

6 |
Jester 2.0: Evaluation of a new linear time collaborative filtering algorithm
- Gupta, Digiovanni, et al.
- 1999
(Show Context)
Citation Context ...borhood methods ignore global relationships between user preferences. In fact global linear relationships between user ratings do exist and were used in the “eigentaste” algorithm by Goldberg et. al. =-=[6]-=-. The eigentaste method is still a neighborhood method, but it uses projections of actual user ratings into a low-dimensional space. This space is computed with a singular-value decomposition of the r... |

1 |
Verifiable secret-sharing and multiparty protocols with honest majority
- Rabin, Ben-Or
- 1989
(Show Context)
Citation Context ...fication is not included in it). We assume that a fraction α > 1/2 of the users’ machines are uncorrupted. Our protocol proceeds in rounds, and we model corruption as a static adversary on each round =-=[1, 12]-=-. That is, the adversary can choose which machines to corrupt at the start of a round, but this choice stays fixed throughout the round. We believe this model is realistic. First of all, in our settin... |