Searchable encryption allows a

Searchable encryption allows auser to send her encrypteddatabase to a server while stillbeing able to perform securesearches over it.

Efficient searchable encryptionschemes leak the search pattern(which can be used to compute thefrequency of each query token) andthe access pattern (which revealshow many documents match thequery).

We propose a new attack (SAP)that uses both search and accesspattern leakage, as well asauxiliary information (which isnot necessarily ground-truthinformation).

To derive the attack, we build amathematical model of theresponse volume and frequencyof each query, based on theauxiliary information.

The attack finds the maximumlikelihood matching of keywordsto query tokens given theprevious mathematical model. Ituses the Hungarian Algorithm tofind the optimal matching.

CLRZ: Chen et al. “Differentially private access patterns for searchable symmetric encryption”. INFOCOM’18PPYY: Patel et al. “Mitigating leakage in secure cloud-hosted data structures”. CCS’19SEAL: Demertzis et al. “SEAL: Attack mitigation for encrypted databases via adjustable leakage”. USENIX’20

SAP is easy to adapt againstdifferent volume-hiding defenses(padding) by just taking thedefense into account in themathematical model.

We evaluate SAP using realdatasets (Enron and Lucene) anduse query frequencies grabbedfrom Google Trends. We give theadversary imperfect auxiliaryinformation to run the attack.

freq: Liu et al. “Search pattern leakage in searchable encryption: Attacks and new construction”. Information Sciences, 2014.graphm: Pouliot and Wright. “The shadow nemesis: Inference attacks on efficiently deployable, efficiently searchable encryption”. CCS’16

By combining both volume andfrequency information, SAPachieves high query recovery(left).SAP outperforms current state-of-the-art attacks (right).

By adapting SAP against previousdefenses, we are able topractically bypass two of them(CLRZ and PPYY) and we stillachieve non-trivial recovery ratesfor the third one (SEAL).

Our results show the importanceof hiding search patterns and/orfrequency leakage. Recent worksthat are moving in this directionseem promising.