Introduction
Deep Neural Networks (DNNs) have achieved remarkable success in computer vision but remain fundamentally fragile. Since the discovery of "adversarial examples" in 2014—inputs with imperceptible changes that trigger incorrect predictions—research has largely treated these as offensive security threats. This survey redefines the field by presenting adversarial attacks as a "dual-use" technology, bridging the gap between their role as system vulnerabilities and their emerging utility as defensive tools for privacy and intellectual property protection.
Taxonomy of Adversarial Attacks
The technical evolution of adversarial attacks can be categorized into three primary domains based on their operational space and objectives.
Pixel-Space Attacks (PSAs)
Pixel-space attacks represent the foundational methods in the field, modifying raw pixel values to induce misclassification. Early methods like the Fast Gradient Sign Method (FGSM) utilized a single-step gradient update for speed. Subsequent developments introduced iterative approaches like Projected Gradient Descent (PGD) and momentum-based variants like MI-FGSM to improve attack success and transferability across different models.
Physically Realizable Attacks
Moving beyond digital manipulation, physical attacks target sensors in the real world. Techniques such as Adversarial Patches, 3D-printed textures, and dynamic light projections (e.g., the SLAP method) demonstrate that DNNs can be fooled by environmental perturbations. Modern "Naturalistic Physical Adversarial Patches" (NPAP) have achieved over 80% success rates against robust detectors like YOLOv3, posing significant risks to autonomous vehicles and surveillance systems.
Latent-Space Attacks (LSAs)
The current frontier involves targeting high-level semantic representations or embeddings. By manipulating the latent manifold of generative models (such as VAEs or Diffusion models), attackers can create "unrestricted" adversarial examples. These images appear perfectly natural to humans but are mathematically "poisoned" for target AI architectures, often exhibiting superior transferability in black-box scenarios.
Adversarial Attacks as a Defensive Tool
A pivotal shift in recent literature is the application of adversarial noise for protection against malicious AI exploitation.
- Privacy Protection: Adversarial perturbations can "immunize" personal photos against unauthorized facial recognition or biometric harvesting.
- Intellectual Property: Tools like Anti-DreamBooth and Unlearnable Diffusion Perturbation (UDP) allow artists to add imperceptible noise to their portfolios. This prevents generative models from effectively training on their data, thereby protecting unique artistic styles from unauthorized mimicry.
Standard Benchmarks and Evaluation
Robustness is typically evaluated using standard computer vision datasets. ImageNet remains the primary benchmark for classification, while COCO and PASCAL VOC are used for object detection and segmentation. In the defensive context, face research datasets and generative model benchmarks are increasingly used to test the efficacy of "unlearnable" perturbations.
Key Insights and Future Directions
The survey highlights a transition from simple sign-based gradients toward more precise Raw Gradient Descent (RGD). While defensive applications show promise, significant challenges remain. "Adversarial purification"—where models attempt to "wash away" noise—continues to threaten the persistence of defensive perturbations. Furthermore, the rise of massive Foundation Models (e.g., CLIP) necessitates new frameworks for understanding vulnerability. Future research must address the computational efficiency of these methods to enable real-time deployment on consumer platforms and social media.
Summary updated