Abstract
Although Deep Neural Networks (DNNs) have been widely applied in various real-world scenarios, they are vulnerable to adversarial examples. The current adversarial attacks in computer vision can be divided into digital attacks and physical attacks according to their different attack forms. Compared with digital attacks, which generate perturbations in the digital pixels, physical attacks are more practical in the real world. Owing to the serious security problem caused by physically adversarial examples, many works have been proposed to evaluate the physically adversarial robustness of DNNs in the past years. In this paper, we summarize a survey versus the current physically adversarial attacks and physically adversarial defenses in computer vision. To establish a taxonomy, we organize the current physical attacks from attack tasks, attack forms, and attack methods, respectively. Thus, readers can have a systematic knowledge of this topic from different aspects. For the physical defenses, we establish the taxonomy from pre-processing, in-processing, and post-processing for the DNN models to achieve full coverage of the adversarial defenses. Based on the above survey, we finally discuss the challenges of this research field and further outlook on the future direction.
Introduction: The Shift to Physical Vulnerabilities
While researchers have long known that Deep Neural Networks (DNNs) can be fooled by "adversarial examples," most early research focused on digital attacks—modifying pixels in a way that is invisible to humans but catastrophic for models. However, digital attacks are rarely practical in the real world because they don't account for the "imaging pipeline," which includes camera lenses, sensors, lighting, and distance.
This survey addresses the gap in physical adversarial attacks, where an attacker modifies a physical object—such as placing a sticker on a stop sign or wearing adversarial glasses—to deceive a computer vision system. These attacks are significantly harder to execute because they must survive environmental changes, yet they pose a much more significant security risk to safety-critical systems like self-driving cars and biometric security.
A Comprehensive Taxonomy of Attacks and Defenses
The primary contribution of this survey is a multi-dimensional taxonomy that organizes over 150 studies into a systematic framework. This allows researchers to categorize threats and defenses based on application, method, and processing stage.
Attack Taxonomy
- Tasks: Categorized by application, including Face Recognition, Self-driving (traffic signs, LiDAR, depth estimation), Security Surveillance (person detection, tracking), and Remote Sensing.
- Forms: How the attack is manifested—Patches (stickers/clothing), Optical (lasers/projections), or 3D-printed objects.
- Methods: Based on the attacker's knowledge—White-box (full access to the model), Black-box (no access, relying on queries or transferability), and Gray-box (partial knowledge).
Defense Taxonomy
- Pre-processing: Detecting and removing patches or smoothing images before they reach the model (e.g., Randomized Smoothing).
- In-processing: Hardening the model itself through Adversarial Training, architecture modifications (like narrowing receptive fields), or Certified Robustness (mathematical guarantees of safety).
- Post-processing: Adding a "re-verification" step where the model's output is checked against logical domain knowledge.
Specialized Benchmarks and Datasets
The paper reviews several specialized benchmarks used to evaluate physical robustness, moving beyond standard digital datasets:
- APRICOT: Over 1,000 "in-the-wild" images of printed adversarial patches.
- 3D Corruptions: Benchmarks for evaluating object detectors against 27 realistic environmental corruptions.
- PAN (Physical Attack Naturalness): Incorporates human ratings to assess how "stealthy" an attack looks.
- PADetBench: A large-scale simulation-based platform for evaluating up to 20 different attack methods against 48 different detectors.
Key Insights and Future Outlook
Physical attacks fail if they don't account for the Image Signal Processor (ISP). Modern attacks have evolved from "meaningless" colorful noise to "meaningful" patches—like cartoon stickers or specific clothing textures—that look benign to humans but are effective against AI. Furthermore, LiDAR is becoming a new frontier, with research into 3D-printed adversarial meshes and laser spoofing.
Despite the rapid evolution of the field, several hurdles remain. Transferability is a major challenge; attacks crafted for one model often fail on another. Additionally, many defenses are too computationally expensive for real-time use on "edge" devices like vehicle computers. The authors argue for the development of Universal Defenses that are model-agnostic and modality-invariant to provide broader protection against the diverse array of physical threats.
Summary updated