Course contentsShow
← Machine Learning and Deep Learning
Lesson 3399 of 3,53874. Adversarial RobustnessFree lesson

Adversarial Attacks on Object Detection

Explore attacks specific to detection systems, including disappearance attacks and mislocalization attacks.

Adversarial Attacks on Object Detection

What you'll learn: How adversarial attacks exploit object detection systems differently than image classifiers, focusing on making objects disappear or appear in wrong locations.

Why Detection Is Different

While image classifiers just output a single label, object detection systems must locate and classify multiple objects simultaneously. This creates unique vulnerabilities. An adversary doesn't just want misclassification—they might want to make a stop sign invisible to a self-driving car or trick a security camera into looking at the wrong spot.

Two Key Attack Types

Disappearance Attacks

These attacks craft perturbations that cause the detector to completely miss an object. Imagine carefully placing stickers on a stop sign that make autonomous vehicle vision systems ignore it entirely—the bounding box never appears. The attack suppresses the detector's confidence below the detection threshold, making the object "vanish."

This is more dangerous than misclassification because the system doesn't even know something is there to be wrong about.

Mislocalization Attacks

These attacks manipulate the bounding box coordinates, causing the detector to identify an object in the wrong location. For example, making a detector think a person is 3 feet to the left of their actual position. This preserves detection but corrupts spatial understanding—critical in robotics, surveillance, or navigation systems.

The Compound Challenge

Object detectors are vulnerable at multiple stages: the feature extraction backbone, the region proposal network, and the classification/localization heads. Adversaries can target any of these components, making defense more complex than single-output classifiers.

Key Takeaway: Object detection systems face unique adversarial threats beyond misclassification—attackers can make objects disappear entirely (disappearance attacks) or appear in incorrect locations (mislocalization attacks), posing serious risks in safety-critical applications like autonomous driving and security systems.