Real-World Adversarial Attacks on AI Systems
Trisha Bera, Sanket Dan, Uddalak Mitra, Jayshree Bhattacharya · 2025
The integration of Artificial Intelligence (AI) into real-world applications has brought societal benefits but also introduced new vulnerabilities, including adversarial attacks. These attacks account for complex physical variables, making them harder to execute and more dangerous. This paper investigates the nature and impact of real-world adversarial attacks on AI systems, categorizing them based on modalities, attack goals, and delivery mechanisms. It critically assesses existing defense mechanisms and suggests emerging trends like certified defenses, sensor fusion techniques, and context-aware learning. The study emphasizes the need for a security-first approach in AI system design and deployment.