Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Survey on Safe Multi-Modal Learning System

Tianyi Zhao, Liangliang Zhang, Yao Ma, Lu Cheng

TL;DR

This work addresses the safety of multimodal learning systems (MMLS) by introducing the first taxonomy organized around four pillars—robustness, alignment, monitoring, and controllability—and applying it to survey methodologies, benchmarks, and open challenges across MMLS. It synthesizes advances in robustness (distributional shift and adversarial attacks), alignment (misalignment/jailbreaking and alignment techniques), monitoring (anomaly detection and reliability of outputs), and controllability (explainability MXAI, fairness, and privacy). The paper identifies gaps such as limited cross-modal defenses, scarce multimodal fairness datasets, and nascent privacy benchmarks for MMLS, and discusses potential directions including standardized datasets, tailored calibration/uncertainty methods, and multimodal-specific privacy and unlearning frameworks. The insights aim to guide future research toward safer deployment of MMLS in high-stakes domains like healthcare and autonomous systems, emphasizing practical benchmarks and robust safety protocols to mitigate risks inherent in cross-modal reasoning and data memorization.

Abstract

In the rapidly evolving landscape of artificial intelligence, multimodal learning systems (MMLS) have gained traction for their ability to process and integrate information from diverse modality inputs. Their expanding use in vital sectors such as healthcare has made safety assurance a critical concern. However, the absence of systematic research into their safety is a significant barrier to progress in this field. To bridge the gap, we present the first taxonomy that systematically categorizes and assesses MMLS safety. This taxonomy is structured around four fundamental pillars that are critical to ensuring the safety of MMLS: robustness, alignment, monitoring, and controllability. Leveraging this taxonomy, we review existing methodologies, benchmarks, and the current state of research, while also pinpointing the principal limitations and gaps in knowledge. Finally, we discuss unique challenges in MMLS safety. In illuminating these challenges, we aim to pave the way for future research, proposing potential directions that could lead to significant advancements in the safety protocols of MMLS.

A Survey on Safe Multi-Modal Learning System

TL;DR

This work addresses the safety of multimodal learning systems (MMLS) by introducing the first taxonomy organized around four pillars—robustness, alignment, monitoring, and controllability—and applying it to survey methodologies, benchmarks, and open challenges across MMLS. It synthesizes advances in robustness (distributional shift and adversarial attacks), alignment (misalignment/jailbreaking and alignment techniques), monitoring (anomaly detection and reliability of outputs), and controllability (explainability MXAI, fairness, and privacy). The paper identifies gaps such as limited cross-modal defenses, scarce multimodal fairness datasets, and nascent privacy benchmarks for MMLS, and discusses potential directions including standardized datasets, tailored calibration/uncertainty methods, and multimodal-specific privacy and unlearning frameworks. The insights aim to guide future research toward safer deployment of MMLS in high-stakes domains like healthcare and autonomous systems, emphasizing practical benchmarks and robust safety protocols to mitigate risks inherent in cross-modal reasoning and data memorization.

Abstract

In the rapidly evolving landscape of artificial intelligence, multimodal learning systems (MMLS) have gained traction for their ability to process and integrate information from diverse modality inputs. Their expanding use in vital sectors such as healthcare has made safety assurance a critical concern. However, the absence of systematic research into their safety is a significant barrier to progress in this field. To bridge the gap, we present the first taxonomy that systematically categorizes and assesses MMLS safety. This taxonomy is structured around four fundamental pillars that are critical to ensuring the safety of MMLS: robustness, alignment, monitoring, and controllability. Leveraging this taxonomy, we review existing methodologies, benchmarks, and the current state of research, while also pinpointing the principal limitations and gaps in knowledge. Finally, we discuss unique challenges in MMLS safety. In illuminating these challenges, we aim to pave the way for future research, proposing potential directions that could lead to significant advancements in the safety protocols of MMLS.
Paper Structure (49 sections, 2 equations, 1 figure)

This paper contains 49 sections, 2 equations, 1 figure.

Table of Contents

  1. Introduction
  2. Robustness
  3. Robustness against Distributional Shift
  4. Understanding distributional robustness in MMLS
  5. Improving MMLS distributional robustness
  6. Adversarial Robustness
  7. Multimodal adversarial attack.
  8. Enhance adversarial robustness of multimodal learning models.
  9. Benchmark
  10. Discussion
  11. Alignment
  12. Misalignment: Jailbreaking MMLS
  13. Aligning MMLS
  14. Benchmark
  15. Discussion
  16. ...and 34 more sections

Figures (1)

  • Figure 1: Taxonomy for Safety of Multimodal Learning Systems.