Through the Lens of LLMs: Unveiling Differential Privacy Challenges
summary
LLMs are used to simulate membership inference and reidentification attacks against Google’s Topics API, finding that a hierarchical model identifies noisy topics with 95.41% accuracy and achieves up to 68.9% reidentification on simulated universal-advertiser scenarios. Embedding analysis with DBSCAN clustering and GloVe embeddings highlights niche or isolated topics within the 349-topic taxonomy that amplify fingerprinting risk, evaluated on a 2,000-user German browsing dataset collected for the experiments.
AI-written summary · about
chapters (8)
- 0:00LLMs and differential privacy
- 1:11LLMs as attack vectors
- 2:14Topics API system overview
- 3:40Membership inference attacks
- 5:27Reidentification attacks
- 6:57Niche topics and edge cases
- 8:42Dataset and limitations
- 9:26Future work and scalability
related talks
35mLeveraging LLM for text augmentation in Topic Modeling
28mAnonymization in the LLM Era: No Room for Compromise
29mAutomating Privacy Reviews with LLMs: Challenges & Opportunities
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
06mData Privacy Concerns and Solutions When using LLMs in Production
33mStargazing into AI Usage with Differential Privacy
10mWhen Privacy Guarantees Meet Pre-Trained LLMs: A Case Study in Synthetic Data
52mAnatomy of an LLM