Demystifying Adversarial Patches: A Threat to Computer Vision

Adversarial Patches

In the ever-evolving landscape of artificial intelligence and machine learning, the term “adversarial patch” has gained significant attention. This technique has been devised to fool machine learning models, particularly those in the realm of computer vision. Adversarial patches can be physical obstructions in captured photos or random alterations applied using algorithms. In this article, we’ll delve into the world of adversarial patches, explore how they can be used, and discuss methods to defend against them.

Understanding Adversarial Patches

Computer vision models are typically trained on straightforward images. These images vary in orientation and resolution but rarely contain patches or unidentified objects. Adversarial patch attacks represent a practical threat to real-world computer vision systems.

How Models Can Be Fooled

Researchers, led by Tom Brown et al., have demonstrated that by placing a digital sticker next to an object in an image, machine learning models can be misled. For example, a banana can be misclassified as a toaster. Experiments conducted by Google’s researchers have paved the way for more systematic methods of generating such patches.

These adversarial patches have the potential to disrupt facial recognition systems, surveillance systems, and even pose challenges to self-driving cars. Besides adversarial patches, there’s a concept called adversarial reprogramming. In this type of attack, a model is repurposed to perform a new task by introducing new parameters into a convolutional neural network. The attacker can attempt to reprogram the network across tasks with significantly different datasets.

Even human-in-the-loop solutions may struggle to identify the intent behind something as ambiguous as a digital sticker.

Is There a Way Out?

Most defenses against patch attacks focus on preprocessing input images to mitigate adversarial noise. What makes this attack significant is that the attacker doesn’t need to know the specific image they are targeting during the attack construction. After generating an adversarial patch, it can be widely distributed for other attackers to use. Existing defense techniques, primarily aimed at small perturbations, may not be robust against larger perturbations.

In a paper under review at ICLR 2020, unnamed authors proposed certified defenses against adversarial patches. They also choreographed white-box attacks to test the model’s resilience further. Additionally, they presented a solution to maintain model accuracy.

Before this work, there were two other approaches aimed at countering adversarial patches:

1. Digital Watermarking (DW)

Hayes in 2018 introduced digital watermarking as a method to detect unusually dense regions of large gradient entries using saliency maps. While this approach led to a 12% drop in accuracy on clean images, it achieved an empirical adversarial accuracy of 63% against non-targeted patch attacks.

2. Local Gradient Smoothing (LGS)

Naseer et al. in 2019 proposed LGS, which is based on the observation that pixel values change sharply within adversarial patches.

Common classification benchmarks often lack inherent protections against adversarial attacks. Researchers at Open AI have introduced a new metric called UAR (Unforeseen Attack Robustness) to evaluate a model’s robustness against unanticipated attacks.

A Proactive Approach

In practice, adversarial attacks can deviate from textbook cases. It’s crucial for machine learning practitioners to identify blind spots within these systems proactively. By designing attacks that expose flaws, developers can better prepare their models for a more diverse range of unforeseen challenges.


Adversarial patches represent a significant challenge to the world of computer vision. These subtle manipulations can fool even the most advanced machine learning models. As the field evolves, so do the methods to defend against such attacks. Understanding the nuances of adversarial patches and staying proactive in defense strategies are crucial in this ever-changing landscape of artificial intelligence.