A Safe, Secure, and Predictable Software Architecture for Deep Learning in Safety-Critical Systems

Biondi, Alessandro; Nesti, Federico; Cicero, Giorgiomaria; Casini, Daniel; Buttazzo, Giorgio

doi:10.1109/les.2019.2953253

Cited by 39 publications

(20 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Experimental results show that real-world applications can benefit from time-predictable FPGA-based hardware acceleration in a GNU/Linux feature-rich operating system environment. Future work should investigate the possibility of integrating the proposed implementation on top of a hypervisor with strong isolation capabilities, providing a safe platform for developing Cyber-Physical Systems [47]. Further developments include improving integration with upcoming releases of Xilinx's FPGA manager driver and minimizing kernel-level code by moving memory buffers allocation and management in userspace.…”

Section: Discussionmentioning

confidence: 99%

A Linux-based support for developing real-time applications on heterogeneous platforms with dynamic FPGA reconfiguration

Pagani

Biondi

Marinoni

et al. 2022

Future Generation Computer Systems

Self Cite

View full text Add to dashboard Cite

Computing platforms for next-generation cyber-physical systems are evolving towards heterogeneous architectures comprising different processing elements and hardware accelerators. In particular, SoC-FPGA platforms, including multiple general-purpose processing cores tightly coupled with an FPGA fabric, represent an attractive solution due to their flexibility, efficiency, and timing predictability. On these platforms, dedicated hardware accelerators implemented on the FPGA fabric can offload computationally intensive activities from general-purpose processing cores. Furthermore, dynamic partial reconfiguration allows virtualizing the FPGA resources by sharing them among multiple hardware accelerators over time.Although very promising, FPGA-based hardware acceleration also introduces new challenges, such as managing and scheduling multiple concurrent acceleration and reconfiguration requests. The FRED framework has been proposed to address these challenges while preserving the predictability required by real-time systems. FRED is based on a device model that matches the capabilities of contemporary SoC-FPGA platforms and comes with an ad-hoc scheduling infrastructure designed to guarantee bounded response times for DPR-enabled accelerated tasks. This paper presents Fred-Linux, the reference implementation of the FRED framework for GNU/Linux. Fred-Linux allows developing rich applications while leveraging predictable FPGA-based hardware acceleration for performing heavy computations. Fred-Linux has been developed using the Zynq-7000 and Zynq-UltraScale+ by Xilinx as reference platforms, and it can be easily ported and extended on other platforms thanks to its modular design.

show abstract

Section: Discussionmentioning

confidence: 99%

A Linux-based support for developing real-time applications on heterogeneous platforms with dynamic FPGA reconfiguration

Pagani

Biondi

Marinoni

et al. 2022

Future Generation Computer Systems

Self Cite

View full text Add to dashboard Cite

show abstract

“…A cross analysis brings a higher chance of catching faulty behavior early. -N-versioning [38]: Running multiple semantically equivalent programs can help finding the cause of the issue. Moreover, since multiple frameworks exist, it is also possible to cross analyze behaviors of different models under different frameworks to detect failure from one of them.…”

Section: Potential Triage Of Silent Bugsmentioning

confidence: 99%

Silent Bugs in Deep Learning Frameworks: An Empirical Study of Keras and TensorFlow

Tambon¹,

Nikanjam²,

An³

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep Learning (DL) frameworks are now widely used, simplifying the creation of complex models as well as their integration to various applications even to non DL experts. However, like any other programs, they are prone to bugs. This paper deals with the subcategory of bugs named silent bugs: they lead to wrong behavior but they do not cause system crashes or hangs, nor show an error message to the user. Such bugs are even more dangerous in DL applications and frameworks due to the "black-box" and stochastic nature of the systems (the end user can not understand how the model makes decisions). This paper presents the first empirical study of Keras and TensorFlow silent bugs, and their impact on users' programs. We extracted closed issues related to Keras from the TensorFlow GitHub repository. Out of the 1,168 issues that we gathered, 77 were reproducible silent bugs affecting users' programs. We categorized the bugs based on the effects on the users' programs and the components where the issues occurred, using information from the issue reports. We then derived a threat level for each of the issues, based on the impact they had on the users' programs. To assess the relevance of identified categories and the impact scale, we conducted an online survey with 103 DL developers. The participants generally agreed with the significant impact of silent bugs in DL libraries and acknowledged our findings (i.e., categories of silent bugs and the proposed impact scale). Finally, leveraging our analysis, we provide a set of guidelines to facilitate safeguarding against such bugs in DL frameworks.

show abstract

“…We here consider an embedded general purpose CPU commonly employed in mission-critical embedded systems, such as space missions (e.g., [25]). The execution model considers a hypervisor for the isolated execution of each filter in a time-triggered fashion, so that misbehaviors are not propagated [26]. Nonetheless, since the proposed approach works at the application level, it can be employed also on other processing platforms, such as GPUs and FPGAs; future work will explore these directions.…”

Section: The Application Model and Hw Platformmentioning

confidence: 99%

Approximation-based Fault Tolerance in Image Processing Applications

Biasielli¹,

Bolchini

Cassano

et al. 2021

IEEE Trans. Emerg. Topics Comput.

View full text Add to dashboard Cite

Image processing applications exhibit an intrinsic degree of fault tolerance due to i) the redundant nature of images, and ii) the possible ability of the consumers of the application output to effectively carry out their task even when it is slightly corrupted. In this application scenario the classical Duplication with Comparison (DWC) scheme, that rejects images (and requires re-executions) when the two replicas' outputs differ in a per-pixel comparison, may be over-conservative. In this paper, we propose a novel lightweight fault tolerant scheme specifically tailored for image processing applications. The proposed scheme enhances the state-of-the-art by: i) improving the DWC scheme by replacing one of the two exact replicas with an approximated counterpart, and ii) allowing to distinguish between usable and unusable images instead of corrupted and uncorrupted ones by means of a Convolutional Neural Network-based checker.To tune the proposed scheme we introduce a specific design methodology that optimizes both execution time and fault detection capability of the hardened system. We report the results of the application of the proposed approach on two case studies; our proposal achieves an average execution time reduction larger than 30% w.r.t. the DWC with re-execution, and less than 4% misclassified unusable images.

show abstract

A Safe, Secure, and Predictable Software Architecture for Deep Learning in Safety-Critical Systems

Cited by 39 publications

References 14 publications

A Linux-based support for developing real-time applications on heterogeneous platforms with dynamic FPGA reconfiguration

A Linux-based support for developing real-time applications on heterogeneous platforms with dynamic FPGA reconfiguration

Silent Bugs in Deep Learning Frameworks: An Empirical Study of Keras and TensorFlow

Approximation-based Fault Tolerance in Image Processing Applications

Contact Info

Product

Resources

About