Hidden item details can be revealed in recommendation systems

Language embeddings used to represent item content may expose item identities to attackers.

Language embeddings used in recommenders can reveal more information than they should. These embeddings may expose item identities to an adversary.

An adversary uses auxiliary item text from external sources to perform an attack. This attack uses cosine-similarity to recover item identities with high mean reciprocal rank.

Language embeddings are used to represent item content. An adversary is someone with access to auxiliary item text. Semantic alignment occurs between independently sourced descriptions.

Leakage is strongest when partial information overlaps with the victim representation. Item leakage can also propagate to user re-identification from sparse ratings.

Researchers evaluated obfuscated descriptions as a mitigation. They found a clear leakage-utility tradeoff.

Key facts

  • A cosine-similarity attack can recover item identities across language models and candidate-set settings.
  • Item leakage can propagate to user re-identification from sparse ratings.
  • Obfuscated descriptions show a clear leakage-utility tradeoff.

All sections

Edition