|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Privacy in User Centric Applications
Hybrid Recommendations with Dynamic Similarity Measure
Participants : Anne-Marie Kermarrec, Nupur Mittal.
This project aims to combine the classical methods of content based and collaborative filtering recommendations, in addition to dynamic similarity computations. The objective is to exploit the varied item-data available from the world wide web, to overcome trivial problems like that of cold-start. In this work, we have designed a new similarity metric inspired from the existing DICE similarity that takes into account changing item/user behavior to compute updated similarity values for the purpose of recommendations. The work leverages the idea of content based recommendations as a first step to create vivid user and item profiles that are iteratively updated.
This work was done in collaboration with Rachid Guerraoui (EPFL, Switzerland), Rhicheek Patra (EPFL, Switzerland).
Lightweight Privacy-Preserving Averaging for the Internet of Things
Participants : Davide Frey, George Giakkoupis, Julien Lepiller.
The number of connected devices is growing continuously, and so is their presence into our everyday lives. From GPS-enabled fitness trackers, to smart fridges that tell us what we need to buy at the grocery store, connected devices—things—have the potential to collect and make available significant amounts of information. On the one hand, this information may provide useful services to users, and constitute a statistical gold mine. On the other, its availability poses serious privacy threats for users. In this work, we designed two new protocols that make it possible to aggregate personal information collected by smart devices in the form of an average, while preventing attackers from learning the details of the non-aggregated data. The first protocol exploits randomness and decomposition into shares as techniques to obfuscate the value associated with each node and lightweight encryption techniques to withstand eavesdropping attacks. The second exploits only randomness and encryption. We carried out a preliminary evaluation and published the results related to the first protocol in [18].
This work was done in collaboration with Tristan Allard from the DRUID Team at IRISA, Rennes.
Collaborative Filtering Under a Sybil Attack: Similarity Metrics do Matter!
Participants : Davide Frey, Anne-Marie Kermarrec, Antoine Rault, Florestan de Moor.
Whether we are shopping for an interesting book or selecting a movie to watch, the chances are that a recommendation system will help us decide what we want. Recommendation systems collect information about our own preferences, compare them to those of other users, and provide us with suggestions on a variety of topics. But is the information gathered by a recommendation system safe from potential attackers, be them other users, or companies that access the recommendation system? And above all, can service providers protect this information while still providing effective recommendations? In this work, we analyze the effect of Sybil attacks on collaborative-filtering recommendation systems, and discuss the impact of different similarity metrics in the trade-off between recommendation quality and privacy. Our results, on a state-of-the-art recommendation framework and on real datasets show that existing similarity metrics exhibit a wide range of behaviors in the presence of Sybil attacks. Yet, they are all subject to the same trade off: Sybil resilience for recommendation quality. We therefore propose and evaluate a novel similarity metric that combines the best of both worlds: a low RMSE score with a prediction accuracy for Sybil users of only a few percent. A preliminary version of this work was published at EuroSec 2015 [57]. This year, we significantly extended the work during the summer internship of Florestan De Moor. Specifically, we considered new attacks that specifically target our novel similarity metric and showed that regardless of the attack configuration, our metric can preserve the privacy of users without hampering recommendation quality. A new paper with these new results was submitted to PETS 2017.
Privacy-Preserving Distributed Collaborative Filtering
Participants : Davide Frey, Anne-Marie Kermarrec.
In this work, we propose a new mechanism to preserve privacy while leveraging user profiles in distributed recommender systems. Our mechanism relies on (i) an original obfuscation scheme to hide the exact profiles of users without significantly decreasing their utility, as well as on (ii) a randomized dissemination protocol ensuring differential privacy during the dissemination process.
We compare our mechanism with a non-private as well as with a fully private alternative. We consider a real dataset from a user survey and report on simulations as well as planetlab experiments. We dissect our results in terms of accuracy and privacy trade-offs, bandwith consumption, as well as resilience to a censorship attack. In short, our extensive evaluation shows that our twofold mechanism provides a good trade-off between privacy and accuracy, with little overhead and high resilience.
This work was done with Antoine Boutet and Arnaud Jegou when they were part of the team, and in collaboration with Rachid Guerraoui from EPFL. But the complete results were published this year in [15].