Cross-Task Transferability of Adversarial Perturbations

This direction is merged into another direction: adversarial transferability. Its last version stays readable and citable here.

Version 1 of 158 papers67 conceptsfrozen 2026-10-04

Traditional adversarial attacks are often viewed as "tricking" a specific model into making a specific mistake—changing a "stop sign" label to "speed limit." However, a growing body of research suggests that adversarial vulnerability is not a quirk of individual classifiers, but a systemic weakness in how neural networks represent visual information.

Researchers have discovered that adversarial perturbations can exhibit cross-task transferability: a malicious image designed to fool an image classifier can simultaneously degrade the performance of an object detector, a semantic segmenter, and even a depth estimation model. This phenomenon reveals that diverse computer vision tasks share a common "vulnerability surface" rooted in their underlying feature extraction backbones.

The Backbone Hypothesis

Why does an attack on a classifier break a depth estimator? The intuition lies in the shared heritage of modern vision models. Whether a model is designed to draw boxes around cars or predict the distance to a wall, it almost always begins by extracting a hierarchy of features—edges, textures, and shapes—using a standard backbone like a ResNet or a Vision Transformer.

Muzammal Naseer, Salman Khan, and their colleagues demonstrated in Neural Representation Distortion (NRD) that by maximizing disruptions in these internal feature spaces rather than the final output layer, they could create "task-generalizable" attacks. Because these mid-level features are "generic" across tasks, disrupting them sabotages the entire downstream pipeline, regardless of whether the final "head" is for classification or segmentation.

This suggests that the vulnerability is not in the task-specific logic, but in the shared feature representation. As Cihang Xie and Lingxi Xie showed with the Dense Adversary Generation (DAG) algorithm, noise generated for one "dense" task (like segmentation) transfers seamlessly to others (like detection), proving that these models rely on a shared set of foundational visual primitives.

Attacking the Representation, Not the Label

To build intuition for why this works, consider the "contrast" of an internal feature map. If a neural network sees a clear edge, that edge is represented by high activation in specific neurons. If an attacker can "flatten" these activations, the model becomes effectively blind to that feature.

Yantao Lu and Yunhan Jia formalized this in their work on Dispersion Reduction (DR). Instead of trying to push an image toward a "wrong" class, DR minimizes the standard deviation (dispersion) of internal feature maps. This makes the image appear "featureless" to the AI. Because every vision task requires some features to function, this "feature-erasing" approach is inherently task-agnostic. It allows a single adversarial image to bypass black-box systems like Google Cloud Vision that perform multiple simultaneous tasks.

Breaking the Primitives: Edges and Textures

The most fundamental shared features are low-level geometric primitives. Christian Cosgrove and Alan L. Yuille explored this by attacking edge detection models. They found that perturbations which cause a model to "hallucinate" or ignore edges have a devastating ripple effect on high-level tasks.

If the structural integrity of edges is compromised, the semantic segmentation model cannot find boundaries, and the classifier cannot recognize shapes. This "bottom-up" vulnerability is so fundamental that Qilong Zhang and Lianli Gao found that attacking low-level features is actually more effective for transferring attacks across different data domains (e.g., from natural images to comics) than attacking high-level semantic objects.

The Multi-Task Achilles' Heel

The trend toward "Unified" or "Multi-Task" networks—where a single model performs detection, depth estimation, and segmentation simultaneously—exacerbates this risk. While these architectures are efficient, they create a single point of failure.

In Multi-Task Adversarial Attack (MTA), Pengxin Guo and Yu Zhang demonstrated that a generative framework can leverage the shared knowledge across tasks to craft more potent attacks. Similarly, UniNet researchers used adversarial attacks as a diagnostic tool, revealing that because tasks like depth estimation and object detection share a single encoder, an attack on one task's objective inevitably bleeds into the others.

Beyond Artificial Vision

Perhaps most strikingly, these shared vulnerabilities may extend beyond silicon. Gamaleldin F. Elsayed and Nicolas Papernot showed that adversarial examples can even transfer to human vision. By mimicking the constraints of the human visual system, they created perturbations that biased human classification decisions. This suggests that the "features" being attacked—edges, textures, and spatial relations—are not just artifacts of deep learning, but fundamental components of how any feedforward visual system, biological or artificial, processes the world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_task_transferability_of_adversaria_63dc93a2,
  title = {Cross-Task Transferability of Adversarial Perturbations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-task-transferability-of-adversarial-perturbations/rdr_63dc93a29a7e45629550a58645fdc4ed},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_task_transferability_of_adversaria_63dc93a2_v1,
  title = {Cross-Task Transferability of Adversarial Perturbations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-task-transferability-of-adversarial-perturbations/rdr_63dc93a29a7e45629550a58645fdc4ed/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}