We propose ICE, an efficient and intelligent system that utilizes our dual-algorithm strategy to achieve superior lossless block-level compression results compared to traditional single compressor methods, especially with intermediate compression levels within the supported range of compressors. ICE employs a multi-step architecture to sample, analyze, plan, and execute compression. Unlike many existing solutions, ICE enables the user to set lower bounds (referred to as ‘PIVOT’) on throughput (compression speed) and compression ratio based on real-world constraints. While meeting these bounds, ICE maximizes the user's desired dimension. Our experimental results over several datasets from different domains demonstrate that ICE can improve the total compression throughput and compression ratio of these datasets with an average of 84% and 1.3% respectively, based on the selected PIVOT value.

Existing oblivious systems offer robust security by concealing memory access patterns, but they encounter significant scalability and performance challenges. Recent efforts to enhance the practicality of these systems involve embedding oblivious computation, e.g., oblivious sorting and shuffling, within Trusted Execution Environments (TEEs). For instance, oblivious sort has been heavily utilized: in Oblix (S&P’18), when oblivious indexes are created and accessed; in Snoopy’s high-throughput oblivious key-value (SOSP’21) during initialization and when the input requests are deduplicated and prepared for delivery; in Opaque (NSDI’17) for all the proposed oblivious SQL operators; in the state-of-the-art non-foreign key oblivious join approach (PVLDB’20). Additionally, oblivious sort/shuffle find applications in Signal’s commercial solution for contact discovery, anonymous Google’s Key Transparency, Searchable Encryption, software monitoring, and differentially private federated learning with user privacy.In this work, we address the scalability bottleneck of oblivious sort and shuffle by re-designing these approaches to achieve high efficiency in distributed multi-enclave environments. First, we propose a multi-threaded bitonic sort optimized for the distributed setting, making it the most performant oblivious sort for small number of enclaves (up to 4). For larger numbers of enclaves, we propose a novel oblivious bucket sort, which improves data locality and network consumption and outperforms our optimized distributed bitonic-sort by up to 5-6×. To the best of our knowledge, these are the first distributed oblivious TEE-based sorting solutions. For reference, we are able to sort 2 GiB of data in 1 second and 128 GiB in 53.4 seconds in a multi-enclave test. A fundamental building block of our oblivious bucket-sort is an oblivious shuffle that improves the prior state-of-the-art result (CCS’22) by up to 9.5× in the distributed multi-enclave setting—interestingly it is better by 10% even in the single-enclave/multi-thread setting.

We propose GraphOS, a system that allows a client that owns a graph database to outsource it to an untrusted server for storage and querying. It relies on doubly-oblivious primitives and trusted hardware to achieve a very strong privacy and efficiency notion which we call oblivious graph processing : the server learns nothing besides the number of graph vertexes and edges, and for each query its type and response size. At a technical level, GraphOS stores the graph on a doubly-oblivious data structure , so that all vertex/edge accesses are indistinguishable. For this purpose, we propose Omix++, a novel doubly-oblivious map that outperforms the previous state of the art by up to 34×, and may be of independent interest. Moreover, to avoid any leakage from CPU instruction-fetching during query evaluation, we propose algorithms for four fundamental graph queries (BFS/DFS traversal, minimum spanning tree, and single-source shortest paths) that have a fixed execution trace , i.e., the sequence of executed operations is independent of the input. By combining these techniques, we eliminate all information that a hardware adversary observing the memory access pattern within the protected enclave can infer. We benchmarked GraphOS against the best existing solution, based on oblivious relational DBMS (translating graph queries to relational operators). GraphOS is not only significantly more performant (by up to two orders of magnitude for our tested graphs) but it eliminates leakage related to the graph topology that is practically inherent when a relational DBMS is used unless all operations are "padded" to the worst case.

We study the problem of multi-user dynamic searchable symmetric encryption (DMUSSE) where a data owner stores its encrypted documents on an untrusted remote server and wishes to selectively allow multiple users to access them by issuing keyword search queries. Specifically, we consider the case where some of the users may be corrupted andcolluding with the serverto extract additional information about the dataset (beyond what they have access to). We provide the first formal security definition for the dynamic setting as well as forward and backward privacy definitions. We then propose$\mu$SE, the first provably secure DMUSSE scheme and instantiate it in two versions, one based on oblivious data structures and one based on update queues, with different performance trade-offs. Furthermore, we extend$\mu$SEto support verifiability of results. To achieve this, users need a secure digest initially computed by the data owner and changed after every update. We efficiently accommodate this, without relying on a trusted third party, by adopting a blockchain-based approach for the digests’ dissemination and deploy our schemes over the permissioned Hyperledger Fabric blockchain. We prototype both versions and experimentally evaluate their practical performance, both as stand-alone systems and running on top of Hyperledger Fabric.

We study the problem of dynamic symmetric searchable encryption. In that setting, it is crucial to minimize the information revealed to the server as a result of update operations (insertions and deletions). Two relevant privacy properties have been defined in that context: forward and backward privacy. The first makes it hard for the server to link an update operation with previous queries and has been extensively studied in the literature. The second limits what the server can learn about entries that were deleted from the database, from queries that happen after the deletion. Backward privacy was formally studied only recently (Bost et al., CCS 2017) in a work that introduced a formal definition with three variable types of leakage (Type-I to Type-III ordered from most to least secure), as well as the only existing schemes that satisfy this property. In this work, we introduce three novel constructions that improve previous results in multiple ways. The first scheme achieves Type-II backward privacy and our experimental evaluation shows it has 145-253X faster search computation times than previous constructions with the same leakage. Surprisingly, it is faster even than schemes with Type-III leakage which makes it the most efficient implementation of a forward and backward private scheme so far. The second one has search time that is asymptotically within a polylogarithmic multiplicative factor of the theoretical optimal (i.e., the result size of a search), and it achieves the strongest level of backward privacy (Type-I). All previous Type-I constructions require time that is at least linear in the total number of updates for the requested keywords, even the (arbitrarily many) previously deleted ones. Our final scheme improves upon the second one by reducing the number of roundtrips for a search at the cost of extra leakage (Type-III).