Encoded but Exposed: Language Embeddings can Reveal More Than They Should in Recommender Systems
A cosine-similarity attack recovers recommender item identities from language embeddings and can aid user re-identification.
Researchers study whether language embeddings used as item features in recommender systems reveal item identities to an adversary with auxiliary text. A simple cosine-similarity attack recovers identities with high mean reciprocal rank across language models and candidate-set sizes on two real-world datasets. Leakage is strongest when auxiliary text overlaps the victim representation, but it also persists through semantic alignment of independently sourced descriptions and can propagate to user re-identification from sparse ratings. Obfuscated descriptions mitigate leakage only with a clear utility tradeoff.
- Cosine similarity recovers item identities at high mean reciprocal rank
- Leakage persists across independently sourced item descriptions
- Item leakage can support user re-identification from sparse ratings
- Obfuscated descriptions reduce leakage only by sacrificing utility
From the original article
1 min read at arxiv.orgLanguage embeddings are increasingly used in recommenders to represent item content, but their privacy risks remain poorly understood. We study whether item embeddings can reveal item identities when an adversary has access to auxiliary item text from external sources.
By Linda Beková · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.