Beyond Boundaries: A Comprehensive Survey of Transferable Attacks on AI Systems

Guangjing Wang, Ce Zhou, Yuanda Wang, Bocheng Chen, Hanqing Guo, Qiben YanarXiv 2023 · DOI · arXiv

Introduction: The Black-Box Challenge

As Artificial Intelligence (AI) becomes deeply integrated into critical infrastructure, from autonomous vehicles to medical diagnostics, the security of these systems is paramount. While much of the early research in adversarial machine learning focused on "white-box" scenarios—where an attacker has full access to a model's architecture and parameters—real-world deployments are typically "black boxes." In these scenarios, attackers must rely on the principle of transferability: the ability of an adversarial attack crafted on a known surrogate model to successfully compromise an unknown target system.

Until now, research into transferable attacks has been fragmented across various domains such as computer vision, natural language processing, and hardware security. This paper provides the first unified, comprehensive survey of transferable attacks, introducing a rigorous taxonomy to map out systemic vulnerabilities across the AI landscape.

A Unified Taxonomy of Transferable Attacks

To bring order to the diverse field of adversarial research, the authors propose a unified six-dimensional taxonomy. This framework categorizes attacks based on two primary branches: the Dimension of transfer (where the attack moves) and the Methodology (how the attack is optimized for transferability).

Figure 1: The proposed taxonomy of transferable attacks, categorized by dimension and methodology.

The Six Dimensions of Transferability

The "pathway" of an attack defines its scope and impact. The survey identifies six distinct axes along which transferability occurs, ranging from simple data generalization to complex cross-hardware exploitation.

Figure 2: Definitions and key literature for the six axes of transferability.
  1. Cross-instance: Attacks that generalize from one specific data sample to others within the same dataset (e.g., Universal Adversarial Perturbations).
  2. Cross-domain: Attacks crafted on one dataset (like ImageNet) that remain effective when the target model is trained on a different domain (like medical imaging).
  3. Cross-modality: A burgeoning frontier where attacks jump between data types, such as using an image-based surrogate to attack a video recognition system.
  4. Cross-model: The most common scenario, where an attack transfers between different architectures, such as from a ResNet-based surrogate to a VGG-based target.
  5. Cross-task: Exploiting shared feature representations to move from one task (e.g., image classification) to another (e.g., object detection).
  6. Cross-hardware: Leveraging physical side-channels, such as power consumption or acoustic signals, to execute attacks across different hardware platforms.

Methodologies for Enhancing Transferability

Creating a transferable attack requires balancing the effectiveness against the surrogate model with the ability to generalize to the target. The paper reviews three core strategies used to achieve this:

  • Data Augmentation: By applying transformations like resizing, padding, or using GAN-generated synthetic data, attackers prevent the adversarial example from overfitting to the specific quirks of the surrogate model.
  • Data Disentanglement: This involves separating "style" from "content." By targeting the core semantic features that most models are likely to share, attackers can create more robust perturbations.
  • Optimization Strategies: Advanced techniques, such as momentum-based Fast Gradient Sign Methods (FGSM) or heuristic searches like Particle Swarm Optimization, help find adversarial directions that are less sensitive to specific model boundaries.

Key Insights: Foundation Models and Inductive Biases

A critical finding of the survey is the role of Foundation Models (e.g., CLIP, BERT, or GPT) as a "Transferability Hub." Because these models are widely used as backbones for downstream applications, a single vulnerability or backdoor implanted in a foundation model can propagate to thousands of different systems. This creates a central point of failure for the entire AI ecosystem.

Furthermore, the paper notes that transferability is often a byproduct of Inductive Bias. Even when architectures differ, models trained on similar data tend to learn similar decision boundaries. Attackers exploit these shared mathematical representations to bypass security measures.

Quantitative Impact and Commercial Vulnerabilities

The survey highlights the real-world potency of these attacks with striking results from recent literature:

  • Commercial APIs: The QueryNet attack demonstrated nearly 100% success rates against commercial black-box APIs, such as Baidu EasyDL, by utilizing sophisticated surrogate ensembles.
  • Physical World Security: Cross-modality attacks have achieved a 73.33% success rate against YOLOv3 using a single physical patch, proving that sensor-fusion systems in autonomous vehicles are not inherently safe.
  • NLP Susceptibility: The TextGuise attack achieved over 80% success with minimal perturbations, showing that text-based models are highly vulnerable to style-transferred adversarial examples.

Conclusion and Future Directions

Despite the rapid progress in identifying vulnerabilities, the authors identify significant gaps in current research. Most notably, the field lacks a standardized benchmark suite to evaluate transferability across all six dimensions. Additionally, while digital attacks are highly effective, transferring them to the physical world—where lighting, angles, and distance vary—remains a significant challenge.

The paper concludes by calling for a shift toward "Holistic Defense." Rather than relying on narrow, reactive patches like adversarial training, the AI community must develop layered security models that integrate sensor-level protection, model diversification, and proactive anomaly detection throughout the AI lifecycle.

Summary updated

BibTeX
@misc{wang2023beyond,
  author = {Guangjing Wang and Ce Zhou and Yuanda Wang and Bocheng Chen and Hanqing Guo and Qiben Yan},
  title = {{Beyond Boundaries: A Comprehensive Survey of Transferable Attacks on AI Systems}},
  year = {2023},
  eprint = {2311.11796},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.11796},
}
Figure