ALERT: Machine Learning-Enhanced Risk Estimation for Databases Supporting Encrypted Queries
Longxiang Wang
34th USENIX Security Symposium (USENIX Security '25) · Day 3 · Privacy 4: Privacy-Preserving Computation
The proliferation of cloud computing and outsourced data storage has led to an increased demand for secure data management solutions, even when data resides on untrusted third-party servers. **Dynamic Searchable Symmetric Encryption (DSSE)** schemes represent a critical advancement in this domain, enabling clients to perform search operations on their encrypted data stored remotely without first decrypting it. DSSE supports three fundamental operations: setup (encrypting and uploading data), query (retrieving specific encrypted files), and update (adding or deleting files), distinguishing it from static searchable encryption. While DSSE aims to provide confidentiality, a significant body of research has demonstrated its vulnerability to **leakage attacks (LAs)**, where adversaries passively monitor query interactions to infer sensitive information.
AI review
Solid applied crypto/ML paper that solves a real and underappreciated problem — leakage attacks on DSSE are genuinely dangerous and the performance-security tradeoff is a legitimate pain point practitioners hit. The reformulation from iterative optimization to offline ML inference is a clean contribution, the engineering details (co-occurrence matrix optimization, dynamic clustering) are concrete, and the benchmarks are honest. Not a landmark paper, but it's doing real work.