A Survey of Adversarial Defenses in Vision-based Systems: Categorization, Methods and Challenges

Adversarial attacks pose a serious threat to real-world AI deployment: The survey identifies adversarial attacks as a "major challenge to the trustworthy deployment of machine learning models," particularly in computer vision applications like image classification and object detection, which are increasingly used in safety-critical systems.

No single defense solution works universally: The paper finds that "multiple different approaches to mitigate different kinds of such attacks are available in the literature, each with their own advantages and limitations," meaning organizations cannot rely on one defensive strategy but must understand the tradeoffs between different protection methods.

Attacks can be implemented in practical, real-world scenarios: Beyond digital attacks, the survey notes that "practical attacks include methods to manipulate the physical world," demonstrating that adversarial threats aren't just theoretical concerns but can affect physical systems—for example, fooling autonomous vehicle vision systems with real-world modifications.

Defense effectiveness varies by attack type and context: The research maps "defenses to the types of adversarial attacks and datasets where they are most effective," indicating that practitioners need to carefully match their chosen defense strategy to their specific threat model rather than assuming one approach protects against all attack variations.

Current defenses have significant shortcomings: The survey emphasizes understanding "shortcomings" of available defenses is "necessary for driving the research in this area," suggesting that existing protection methods remain incomplete and organizations deploying vision-based AI systems should be aware of remaining vulnerabilities.