DarkStream: Exploiting Internal Throughput Contention in Data Streaming Accelerator for Timing Attacks

Modern Intel processors integrate a variety of off-core accelerators to enhance data movement efficiency for memory-intensive workloads in data centers. Among them, the Data Streaming Accelerator (DSA) offloads repetitive memory operations such as copying and integrity checking to improve performance and energy efficiency. The DSA employs the group abstraction to isolate each client's work queues, engines, and arbiters, thereby aiming to prevent interference and performance degradation among concurrent tenants. Despite the isolation provided by the group abstraction, our analysis reveals that an I/O fabric interface beneath the group boundary remains shared across all clients in the DSA. As a result, this microarchitectural design can trigger performance contention among clients, leading to timing variations across them. In this paper, we present DarkStream, which exposes vulnerabilities arising from the DSA's internal throughput contention. We demonstrate that these timing differences can be exploited to construct a covert channel with a bandwidth of up to 129 Kbps and to perform website and deep learning model fingerprinting with 97.03% and 99.17% accuracy, respectively.