Real-Time Scheduling of TrustZone-enabled DNN Workloads

Babar, Mohammad Fakhruddin; Hasan, Monowar

doi:10.1145/3560826.3563386

Cited by 2 publications

(6 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A similar line of work exists [38] that presents a technique to enable time-critical DNN computations for TrustZoneenabled enclaves. As TrustZone enclaves do not have enough capacity to process DNN models from multiple tasks and run them together while retaining temporal constraints, the authors proposed splitting the DNN tasks into multiple chunks.…”

Section: ) Time-aware Trusted Dnn Executionmentioning

confidence: 98%

“…The majority of the work (i.e., 11 out of 16) [30], [31], [32], [33], [34], [35], [36], [37], [38], [39], [40] assume blackbox attacks and the rest consider white-box attacks [41], GAN-based attacks [42], fault injection attacks [43], [44], and membership inference scenarios [45]. A typical assumption is that adversaries have complete access to the rich OS.…”

Section: A Threat Models Used In Literaturementioning

confidence: 99%

“…A stealthy attacker could try to access the model's parameters, intermediate results, and final output without disrupting the inference task (T-Slices [40]). Further, the adversary may compromise the rich OS to intrude, forge, and modify the interference task and steal user data from nonprotected processes (SecDeep [31], real-time DNN [38]). They may also have access to the cache, memory, and storage (Trusted-DNN [33]) and inject faults that may change specific DNN parameters in the memory (AegisDNN [43]).…”

Section: A Threat Models Used In Literaturementioning

confidence: 99%

“…Researchers proposed several strategies depending on application requirements. For instance, there exist models for running the entire DNN layers inside TEEs [30], [33], [34], [38], [40]. Another alternative is to run the model in multiple enclaves to speed up the inference process [30], [32], [35], [37], [39], [42], [43], [45].…”

Section: A Summary and Observationsmentioning

confidence: 99%

“…Table 3 summarizes the techniques we discuss in this survey. Majority of work (14 out of 16) [30], [31], [32], [33], [34], [35], [36], [37], [38], [40], [42], [43], [44], [45] focus on inference task and remaining two [39], [41] consider both training and inference (see ''Learning Task'' column). We use the ''Design Focus'' column to demonstrate the key performance metrics considered by the authors.…”

Section: A Summary and Observationsmentioning

confidence: 99%

See 4 more Smart Citations

Trusted Deep Neural Execution—A Survey

Babar

Hasan

2023

IEEE Access

Self Cite

View full text Add to dashboard Cite

The growing use of deep neural networks (DNNs) in various applications has raised concerns about the security and privacy of model parameters and runtime execution. To address these concerns, researchers have proposed using trusted execution environments (TEEs) to build trustworthy neural network execution. This paper comprehensively surveys the literature on trusted neural networks, viz., answering how to efficiently execute neural models inside trusted enclaves. We review the various TEE architectures and techniques employed to achieve secure neural network execution and provide a classification of existing work. Additionally, we discuss the challenges and present a few open issues. We intend that this review will assist researchers and practitioners in understanding the state-of-the-art and identifying research problems.

show abstract

Section: ) Time-aware Trusted Dnn Executionmentioning

confidence: 98%