Cross Model Adversarial Transferability

2,619 concepts1,598 papers75 proposals

Adversarial examples are often perceived as highly specific, brittle glitches tailored to a single neural network instance. However, a remarkable and troubling property of deep learning models challenges this assumption: adversarial transferability. An adversarial perturbation optimized completely offline against a local, accessible "surrogate" model can reliably deceive a remote, proprietary "target" model—even when the attacker has zero knowledge of the target's internal parameters, architecture, or weights.

To understand why this works, we must look at the geometry of high-dimensional decision spaces. When different models are trained on the same dataset or task, they do not discover completely disparate solutions. Instead, they converge on similar representational spaces and carve up their decision boundaries in remarkably consistent ways. An adversarial vector that successfully pushes an input across a surrogate model's decision boundary is highly likely to carry that same input across the parallel decision boundary of an unknown target model.

The Surrogate Framework and Zero-Query Attacks

In a realistic threat model, an adversary interacting with a commercial API or a secure biometric system cannot access the target model's gradients directly. To circumvent this, attackers deploy the surrogate model pipeline. By training a local proxy model to mimic the target domain, the attacker transforms a blind black-box problem into an accessible white-box environment.

[Input Data] ---> [Local Surrogate Model] ---> Compute Gradients ---> [Adversarial Perturbation]
                                                                               |
                                                                       (Transfers to)
                                                                               v
[Adversarial Input] --------------------------------------------> [Remote Target Model] (Fooled!)

This extraction of decision boundaries can be highly efficient. For example, Xuanli He et al. demonstrated in Model Extraction and Adversarial Transferability, Your BERT is Vulnerable! that high-fidelity model extraction can be used to build functional white-box clones of private NLP APIs. Adversarial examples generated on these extracted surrogates maintain their potency when transferred back to the original black-box system, bypassing security measures without requiring direct access to the target parameters.

The true power of transferability lies in its ability to enable zero-query attacks. Traditional black-box attacks rely on query-based optimization, sending thousands of slightly modified inputs to a target system to iteratively estimate its gradients. This approach is slow, computationally expensive, and easily flagged by anomaly detection systems. Transferability eliminates this feedback loop entirely.

In the domain of voice biometrics, Guangke Chen et al. leveraged this in QFA2SR: Query-Free Adversarial Transfer Attacks to Speaker Recognition Systems. By optimizing perturbations on local surrogates specifically for transferability rather than local accuracy, the authors successfully fooled proprietary speaker recognition systems without a single target interaction. Similarly, Zikui Cai et al. showed in Zero-Query Transfer Attacks on Context-Aware Object Detectors that by utilizing pre-computed success probabilities and context graphs, an attacker can execute highly transferable, single-shot adversarial attacks that evade traffic monitoring defenses entirely.

Shared Lineage and Inherited Vulnerabilities

The success rate of a transfer attack is intrinsically tied to the structural and historical relationship between the surrogate and target models. The modern machine learning ecosystem relies heavily on shared foundations—such as public pre-trained models and standardized compression pipelines—which inadvertently creates systemic vulnerabilities.

Consider the common industry practice of deploying lightweight, compressed models to edge devices while keeping the massive parent models secure on a private server. Sanghyun Hong et al. exposed a critical flaw in this pipeline in Publishing Efficient On-device Models Increases Adversarial Vulnerability. Because the public on-device model shares a common lineage and training distribution with its parent, it functions as an ideal white-box surrogate. Attackers can compute precise adversarial gradients on their local mobile devices, generating perturbations that seamlessly transfer to compromise the hidden, high-stakes server-side model.

A parallel vulnerability emerges from transfer learning and fine-tuning. When developers initialize a private target model using a popular open-source backbone, they inherit its geometric properties. Yangqiu Song et al. quantified this phenomenon in Two Sides of the Same Coin: White-box and Black-box Attacks for Transfer Learning. Their work revealed a stark trade-off: while fine-tuning a model on a downstream task can increase its resistance to direct white-box manipulation, it simultaneously leaves the model highly susceptible to black-box transfer attacks launched from its original public source model.

Crossing Structural and Modality Divides

Intuitively, one might expect adversarial transferability to break down when moving between entirely different neural network architectures or data representations. If Model A is a Convolutional Neural Network (CNN) that processes local pixel patches and Model B is a Vision Transformer (ViT) that relies on global self-attention, their internal mechanics are fundamentally distinct. Yet, empirical evidence shows that transferability transcends these structural divides.

This cross-architecture resilience occurs because diverse models still learn to align their internal representations with the same underlying semantic features of the physical world. Kunyang Li et al. explored this in The Efficacy of Transformer-based Adversarial Attacks in Security Domains, finding that ViTs act as highly potent generators of transferable attacks. Perturbations optimized on a ViT's attention maps exhibit high transmissibility, successfully blinding alternative architectures like CNNs. This cross-model vulnerability extends to proprietary, multimodal foundation models; Yichi Zhang et al. demonstrated in How Robust is Google's Bard to Adversarial Image Attacks? that adversarial images optimized on accessible, open-source vision encoders can reliably bypass the safety and identity filters of closed-source systems like Google's Bard.

Even more striking is cross-representation transferability, where an attack survives a complete transformation of the underlying data format. In Cross-Representation Transferability of Adversarial Attacks: From Spectrograms to Audio Waveforms, Sajjad Abdoli demonstrated that adversarial perturbations designed to fool a 2D image classifier processing audio spectrograms can be inverted and reconstructed into a 1D time-domain audio waveform. Despite the mathematical transformation, the resulting raw audio file successfully deceives a completely different 1D neural network that processes raw audio directly.

Turning Vulnerability into a Defensive Shield

While adversarial transferability represents a profound security challenge for AI deployment, it can be inverted to serve as a powerful mechanism for user privacy. In an era where personal data is routinely scraped to train unauthorized surveillance models, individuals can weaponize transferability to proactively protect their digital footprints.

[User Photo] ---> Apply Mask Optimized on Local Surrogate ---> [Protected Photo]
                                                                      |
                                                              (Uploaded to Web)
                                                                      v
                                                       [Unknown Surveillance Model]
                                                       (Fails to recognize identity!)

This defensive application relies on the fact that a protective mask does not need to know the exact architecture of a malicious observer to be effective. In the visual domain, Xiao Yang et al. introduced this paradigm in Towards Face Encryption by Generating Adversarial Identity Masks. By generating subtle, human-imperceptible identity masks on local surrogate models, users can encrypt their photos before uploading them to social media. Because the adversarial noise transfers across architectures, it successfully disrupts downstream, black-box facial recognition systems.

A similar protective barrier can be constructed for biometric audio data. Chien-yu Huang applied proxy-based adversarial transfer to voice privacy in Defending Your Voice: Adversarial Attack on Voice Conversion. By overlaying a calculated, imperceptible noise layer onto voice recordings, the author demonstrated that users can intentionally break unauthorized voice-cloning and synthesis tools.

By exploiting the shared geometric vulnerabilities inherent to deep learning, these defensive frameworks ensure that privacy protection remains robust, even when facing an asymmetric landscape of unknown, evolving target models.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 1597 papers
  2. Version 1 · 2026-10-04 · 1314 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_adversarial_transferability_f71e57b3,
  title = {adversarial transferability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-transferability/rdr_f71e57b3b4f34ed492747219c7cfc398},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_adversarial_transferability_f71e57b3_v2,
  title = {adversarial transferability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-transferability/rdr_f71e57b3b4f34ed492747219c7cfc398/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake