Ghost Recommendations: A Protocol for Efficiently Enhancing User Privacy
Bibliographic record
Abstract
As the amount of online information accessible to users keeps increasing, we have come to rely more on services such as Netflix, Amazon, and eBay that are successful in recommending choices to users. The main goal of such services is to present the user with a more personalized set of choices or recommendations. The growing importance of recommendation systems that provide these services can be attested by the efforts the academic community is taking towards improving their performance. The quality of recommendation systems is primarily determined by the accuracy of the results they can provide to the users. To achieve high-accuracy results, these systems count on finding similarities between different users based on various features, such as the ratings the users provide for the items. Recommendation systems use different techniques, often harvesting private user information, to detect these similarities. Therefore, providing better recommendations frequently comes at the cost of user privacy and at the risk of exposing the user's preferences. Owing to growing concerns about this risk, researchers started to investigate recommendation solutions with better assurances of privacy. \n \nThere is a growing body of work with respect to making recommendation systems more sensitive towards user privacy. The current solutions implemented use various methodologies like randomization of the dataset, anonymizing the identities of users, using data aggregation, obfuscating user data, using a trusted third party, and using cryptographic techniques. However, we are yet to have a solution that not only provides privacy guarantees, but is also a practical and efficient system, giving recommendations with high accuracy. \n \nOur goal in this thesis is to implement a solution that enables high guarantees of user privacy, is practical and efficient, that scales well over a large dataset, and provides users with accurate recommendations. A common trend in the solutions mentioned before is to model a system around one or more trusted third parties. All the critical operations such as key generation or user authentication are delegated to these trusted third parties and combined with a threat model that restricts them from behaving in a malicious manner. We aim at implementing a system that is independent of such a trusted third party. We also desire a system that makes collusion among servers ineffective unless the number of corrupt servers exceeds a threshold value. We also want to make all computations independent of the availability of the participants, so that users would get recommendations even if other participants are offline. For our use case, we have considered a scenario where users would like to get recommendations of movies that are based on ratings provided for other movies. To evaluate our system we have used the real world, publicly available “MovieLens” dataset. Our system consists of the following entities: a set of users or clients, a distributed set of servers, and a public bulletin board. Our scheme primarily focuses on maintaining the privacy of user preferences as well as the recommendations and it does not allow anyone other than the user herself to have access to the data.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.017 | 0.035 |
| Meta-epidemiology (narrow) | 0.001 | 0.002 |
| Meta-epidemiology (broad) | 0.002 | 0.002 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.004 | 0.004 |
| Scholarly communication | 0.005 | 0.013 |
| Open science | 0.005 | 0.013 |
| Research integrity | 0.007 | 0.010 |
| Insufficient payload (model declined to judge) | 0.017 | 0.010 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".