Message from the DSML 2022 Organizers

Homa Alemzadeh, Rakesh B. Bobba, Varun Chandrasekaran, David Evans, Nicolas Papernot, Karthik Pattabiraman, Florian Tramèr2022 · DOI

Introduction

As machine learning (ML) transitions from research environments to high-stakes, safety-critical deployments, a fundamental "reality gap" has emerged. While models often achieve state-of-the-art accuracy on static, labeled datasets, they frequently lack the reliability, security, and formal guarantees required for real-world operation. The 5th Workshop on Dependable and Secure Machine Learning (DSML 2022) was organized to address this gap, serving as a vital bridge between the ML and dependable computing communities.

The Problem: The Fragility of Modern ML

The organizers of DSML 2022 identify a core flaw in the current ML development lifecycle: the assumption that training data is perfectly representative of the real world. In "wild" deployments—such as autonomous vehicles, healthcare systems, and criminal sentencing algorithms—this assumption fails, exposing three primary vulnerabilities:

  1. Unexpected Perturbations: Models are susceptible to both accidental environmental noise and maliciously crafted adversarial attacks that can lead to catastrophic failures.
  2. Over-confidence: Current algorithms often provide high-certainty predictions even when processing out-of-distribution inputs. This lack of self-awareness makes it difficult to implement necessary "failsafe" modes.
  3. Hardware Vulnerabilities: The increasing reliance on specialized hardware accelerators introduces a new attack surface, as these physical components are themselves susceptible to faults and side-channel vulnerabilities.

Community-Driven Solutions

To combat these challenges, DSML 2022 established an interdisciplinary forum for researchers, practitioners, and regulatory experts. The workshop curated high-quality research through a rigorous peer-review process, accepting 7 out of 10 submissions (a 70% acceptance rate), with each paper evaluated by 3-4 experts.

The research agenda was further shaped by two thematic keynotes:

  • Siva Kumar Sastry Hari (NVIDIA): Addressed the safety and resilience of Autonomous Vehicles, focusing on how systems can withstand hardware and software faults.
  • Alfred Chen (UC Irvine): Introduced the concept of "Semantic AI Security" in Cyber-Physical Systems, advocating for a shift from pixel-level analysis to higher-level logical and contextual security.

Key Insights and Strategic Shifts

The workshop highlighted several paradigm shifts necessary for the future of dependable ML:

The Accuracy Trap

High accuracy on a test set is no longer a sufficient metric for success in critical domains. The organizers argue that reliability and robustness must be treated as primary requirements from the design phase, rather than afterthoughts added during deployment.

The Need for Failsafes

A major hurdle in safety-critical settings is the inability of models to revert to a failsafe mode. Future research must focus on enabling models to recognize their own limitations and hand over control when they encounter data they cannot reliably process.

Hardware-Software Co-design

Dependability cannot be achieved through software alone. The workshop explicitly called for hardware-software co-design, acknowledging that the security of ML systems is inextricably linked to the integrity of the underlying hardware accelerators.

Future Outlook and Challenges

Despite the progress made at DSML 2022, significant hurdles remain. The organizers identified three critical areas for future work:

  • Compliance: Determining how complex, non-linear ML models can comply with existing safety-critical system requirements and certifications.
  • Community Silos: Continuing to break down the barriers between the machine learning and dependable computing communities to foster collaborative problem-solving.
  • Scaling Guarantees: As models grow in complexity, the difficulty of establishing formal guarantees of safety grows exponentially. Developing scalable verification and validation methods remains a primary research goal.

Conclusion

DSML 2022 successfully transitioned back to a hybrid format in Baltimore, ensuring global accessibility while fostering in-person collaboration. By focusing on the intersection of theoretical security and practical application, the workshop continues to play a pivotal role in defining the standards for the next generation of secure and dependable AI systems.

Summary updated

BibTeX
@misc{alemzadeh2022message,
  author = {Homa Alemzadeh and Rakesh B. Bobba and Varun Chandrasekaran and David Evans and Nicolas Papernot and Karthik Pattabiraman and Florian Tramèr},
  title = {{Message from the DSML 2022 Organizers}},
  year = {2022},
  doi = {10.1109/dsn-w54100.2022.00008},
}
Figure