Message from the DSML 2019 Organizers

Homa Alemzadeh, Rakesh B. Bobba, David Evans, Nicolas Papernot, Karthik Pattabiraman, Florian Tramèr2019 · DOI

Abstract

The 2019 Workshop on Dependable and Secure Machine Learning (DSML) serves as a pivotal forum for addressing the inherent fragility of modern machine learning (ML) models. While ML has achieved remarkable success in controlled environments, its deployment in safety-critical applications remains hampered by vulnerabilities to adversarial attacks, environmental noise, and hardware faults. This message from the organizers outlines the workshop’s mission to bridge the gap between ML performance and dependable computing, advocating for a shift toward verifiable and resilient systems.

Introduction

The rapid integration of machine learning into critical infrastructure—ranging from autonomous vehicles to healthcare diagnostics—has exposed a significant gap in our engineering practices. Traditionally, the ML community has prioritized predictive accuracy on static, labeled datasets. However, this focus often ignores the "wild" conditions of real-world deployment. The DSML 2019 workshop was established to address this "fragility," challenging researchers to move beyond empirical performance toward the engineering of dependable, secure, and verifiable systems.

The Fragility of Modern Machine Learning

The organizers identify three primary dimensions where current ML models fall short of the requirements for safety-critical deployment:

  1. Input Sensitivity: Models often exhibit extreme sensitivity to small perturbations. Whether caused by accidental noise or intentional adversarial attacks, these minor changes can lead to catastrophically incorrect predictions made with high confidence.
  2. Lack of Failsafes: Unlike traditional safety-critical software, many ML models lack a "failsafe" mode. They do not possess an inherent mechanism to signal uncertainty or revert to a safe state when encountering out-of-distribution data.
  3. Hardware Vulnerabilities: The reliability of ML is inextricably linked to the hardware it runs on. Specialized accelerators are susceptible to physical faults and side-channel attacks, yet these hardware-level risks are rarely integrated into the ML security perimeter.

Workshop Methodology and Curation

To address these challenges, the DSML 2019 workshop curated a program designed to foster interdisciplinary dialogue between the ML and dependable computing communities. The selection process was rigorous, with 17 submissions undergoing peer review by at least three experts each. Ultimately, 10 high-quality contributions (9 papers and 1 talk) were selected, representing a ~58% acceptance rate.

The workshop was structured around three thematic pillars:

  • Industry Perspectives: Focusing on privacy and the practical challenges of deploying ML in commercial environments.
  • Attacks and Defenses: Academic research into the robustness of models against adversarial manipulation.
  • Resilient Systems: The design of fault-tolerant architectures and the integration of formal verification methods.

Key Insights: Beyond Accuracy

A central theme of the workshop is the necessity of a hybrid approach that combines the strengths of machine learning with the rigor of dependable computing. Key insights include:

  • New Metrics: Accuracy is insufficient for safety-critical systems. The community must develop standardized metrics for "accountability" and "robustness."
  • Hardware-Software Co-design: Reliability must be addressed at both the algorithmic and hardware levels. The underlying hardware must be considered part of the security perimeter.
  • Verified AI: Moving toward mathematical guarantees of safety through formal verification is essential for building trust in "black-box" models.

Conclusion and Future Directions

While DSML 2019 successfully established a forum for these critical issues, the field remains in its infancy. Future work must focus on building a cohesive research community that spans multiple disciplines. Furthermore, significant hurdles remain regarding the scalability of formal verification for complex models and the alignment of ML systems with existing legal and safety compliance frameworks. The organizers conclude by congratulating the contributors and inviting the community to continue this essential work toward trustworthy artificial intelligence.

Summary updated

Abstract

Presents the introductory welcome message from the conference proceedings. May include the conference officers' congratulations to all involved with the conference event and publication of the proceedings record.
BibTeX
@misc{alemzadeh2019message,
  author = {Homa Alemzadeh and Rakesh B. Bobba and David Evans and Nicolas Papernot and Karthik Pattabiraman and Florian Tramèr},
  title = {{Message from the DSML 2019 Organizers}},
  year = {2019},
  doi = {10.1109/dsn-w.2019.00007},
}
Figure